Estimation Real Number of Road Accident Casualties

D.1.15. Final Report on Task 1.5 

Contract No: TREN-04-FP6TR-SI2.395465/506723 “SafetyNet” 

Acronym: SafetyNet 

Title: Building the European Road Safety Observatory 

Integrated Project, Thematic Priority 6.2 “Sustainable Surface Transport” 

Project Co-ordinator: 

Professor Pete Thomas 

Vehicle Safety Research Centre 

Ergonomics and Safety Research Institute 

Loughborough University 

Holywell Building 

Holywell Way 

Loughborough 

LE11 3UZ 

Organisation name of lead contractor for this deliverable: TRL 

Due Date of Deliverable: 31/12/2007 

Submission Date: 20/03/2008 

Report Authors: Jeremy Broughton, Emmanuelle Amoros, Niels Bos, 

Petros Evgenikos, Stefan Hoeglinger, Péter Holló, 

Catherine Pérez, Jan Tecl 

Project Start Date: 1st May 2004 

Duration: 4 years 

Project co-funded by the European Commission within the Sixth Framework Programme (2002 -2006) 

Dissemination Level 

PU Public X 

PP 

RE 

CO 

Restricted to other programme participants (inc. Commission Services) 

Restricted to group specified by consortium (inc. Commission Services) 

Confidential only for members of the consortium (inc. Commission Services) 

Project co-financed by the European Commission, Directorate-General Transport and Energy 

sn_trl_1.5_final report_1 02/04/2008

SafetyNet Deliverable D1.15 

Table of Contents 

1. Executive Summary 4 

2. Introduction 8 

2.1 Background 10 

2.2 The common methodology 11 

2.3 Injury and severity coding 14 

3. Results 16 

3.1 Results from the UK study 17 

3.2 Results from the other studies 23 

4. Synthesis 32 

4.1 MAIS and LoS compared 33 

4.2 Definition of ‘hospitalised’ person 36 

4.3 Trends in the Conversion Factors 36 

4.4 Adjustment of ICD10-based results 37 

4.5 An application 40 

5. Conclusions and Recommendations 42 

6. References 45 

7. Appendix A: Reports of the national studies 46 

7.1 Austria 46 

7.1.1 Introduction 46 

7.1.2 Description of data sources 46 

7.1.3 Description of linking process 47 

7.1.4 Results 60 

7.1.5 Conclusions 70 

7.1.6 References 71 

7.2 Czech Republic 72 







sn_trl_1.5_final report_1 02/04/2008 Page 2


7.3 France 80 







7.4 Greece 92 




7.4.4 Results 100 



7.5 Hungary 106 




7.5.4 Results 110 


7.6 Netherlands 117 




7.6.4 Results 143 



7.7 Spain 166 




7.7.4 Results 178 



7.8 United Kingdom 186 




7.8.4 Results 191 






1. Executive Summary 

The objective of Task 1.5 of the SafetyNet IP has been to estimate the actual 

numbers of road accident casualties in Europe from the CARE database by 

addressing two issues: 

• the under-reporting in national accident databases and 

• the differences between countries of the definitions used to classify injury 

severity. 

Currently, the only comparable measurement units available in CARE are the 

numbers of fatal accidents and of people killed, where the degree of underreporting 

is acceptably small in most EU Member States and there is a common 

definition. The same is not true, however, of non-fatal accidents and of 

casualties who are not killed. As a result, at present the numbers of non-fatal 

accidents and of people seriously and slightly injured cannot be compared in 

different Member States. In addition, the definition of injury severity differs 

among member states, so that a casualty which would be recorded in one 

country might not be recorded in another. Equally, a casualty which might be 

recorded as ‘seriously’ injured in one country might be recorded as ‘slightly’ 

injured in another. 

As a result of this lack of comparability, international comparisons of road safety 

focus entirely on fatal accidents and fatalities, which form only a small minority 

of the totals. It is highly desirable to extend these comparisons to include the full 

range of injury severities. The objective of Task 1.5 has been to allow this to 

happen. 

In order to overcome the inconsistencies in the reporting of non-fatal casualties, 

this Task has: 

1. estimated the under-reporting level for non-fatal casualties by developing a 

uniform methodology and applying it in several EU countries, 

2. estimated the number of serious casualties per country according to a new 

common measurement unit. 

This report documents the results that have been achieved: 

° The study began by agreeing a common methodology that would be 

applied by all partners in Task 1.5 for their studies. 

° Studies were carried out in 8 countries according to this methodology, and 

the report contains detailed descriptions of the individual studies. 

° In each study, files of police and hospital records were assembled for the 

road accidents that occurred in a common area. These files were 

compared to identify matching records, i.e. those casualties who were 

present in both files. For these matching records, certain medical details 

were added to the police records: length of stay in hospital and injury 

severity (specifically the Maximum Abbreviated Injury Score MAIS, an 

internationally accepted summary measure of injury severity). 




° Two matrices were then prepared to summarise the outcome of each 

study, one based on injury severity and the other on length of stay. 

° These matrices were brought together for analysis, and conversion factors 

for each study were estimated in a consistent way. These factors allow the 

actual number of serious casualties in each country to be estimated 

consistently from police accident statistics. 

° The new common measurement unit for counting serious casualties could 

be based on either injury severity or length of stay. It is concluded that the 

most robust definition is of a non-fatal casualty with MAIS>=3 (inclusive). 

° Initial comparisons have been made of casualty data adjusted by the 

conversion factors estimated by the national studies. 

The various national studies encountered a range of problems concerning 

access to the hospital data and content of the data. In general these were 

overcome successfully, although there were some implications for the results 

that could finally be achieved. 

The coverage of the studies varied widely, influenced to some extent by 

whether hospital data had to be collected directly (as in the Czech Republic and 

Hungary) or were already available from files that had been compiled by 

national or regional authorities. The size of the datasets varies widely, 

depending on the size of the study area and the period included. The studies 

are summarised below. 

Country Study area Period 

Austria National 2001 

Czech Republic Local (Kromeriz) 2003 - 2005 

France Regional (Département of the Rhône) 1996 - 2003 

Greece Regional (Corfu) 1996 - 2003 

Hungary Local (part of Budapest) Aug 2004 - Jan 2006 

The Netherlands National 1997 - 2003 

Spain Regional (Castilla y Leon) July - Dec 2005 

United Kingdom Regional (Scotland) 1997 - 2005 

Ideally, these studies would have covered complete countries and so been truly 

national. Only 2 studies were truly national, so the question arises in the 

remaining 6 countries of whether conversion factors estimated from subnational 

studies can be generalised to the national data. The answer must vary 

from country to country, but in general the larger the study area the more likely 

the conversion factors are to be nationally representative. 

The new common measurement unit is a non-fatal casualty with MAIS>=3. Most 

of these are recorded by the police as seriously injured, but the studies show 

that the police record some as slightly injured. Consequently, according to this 

definition the number of casualties C in a particular country is estimated as 




C = N1 * police reported serious casualties + 

N2 * police reported slight casualties 

where N1 and N2 vary from country to country. The overall factors from 7 

studies are shown below (they could not be estimated in Austria because of 

data limitations). N2 is considerably smaller than N1 and hence is multiplied by 

10 in this figure. 

Conversion Factors for MAIS>=3, all road users 

Conversion Factors 

1.0 

0.5 

N1 N2*10 

0.0 

France Hungary Greece Netherlands 

Czech 

Republic 

UK 

Spain 

It was originally envisaged that the conversion factors would be generalised to 

other countries, in order to increase the utility of the CARE database. However, 

the results have led to the conclusion that this would not provide reliable results. 

The only satisfactory approach would be to carry out comparable studies in as 

many countries as possible. 

The results from the Dutch and UK studies have also shown that the conversion 

factors can change through time as police accident reporting practices evolve. 

Thus, studies need to be repeated regularly to update the factors. 

In summary, the research that has been carried out in the course of SafetyNet 

Task 1.5 represents a significant step forward and allows for the first time the 

number of severely injured casualties to be compared meaningfully between 

countries. The goals of the research were ambitious, but the practical problems 

that were encountered have meant that some could not be achieved fully. The 

lessons that have been learnt will allow this type of study to be carried out more 

effectively in future. 

The central problem of this type of study is of obtaining access to anonymised 

medical records. Access to these records for research purposes is often 

problematical. Modern linkage techniques such as those used in this study, 

however, make these data increasingly valuable. Ways need to be found to 




persuade the custodians of these data to allow them to be used for purposes 

that support the broader aims and welfare of society. 




2. Introduction 

The objective of Task 1.5 of the SafetyNet IP is to estimate the actual numbers 

of casualties in Europe from the CARE database by addressing the issue of 

under-reporting and the differences in national systems for injury classification. 

Currently, the only comparable measurement units available in CARE are the 

numbers of fatal accidents and of people killed, where the degree of underreporting 

is acceptably small in most EU Member States. The same is not, 

however, true of non-fatal accidents and of injured people who do not die, so at 

present the numbers of non-fatal accidents and of people seriously and slightly 

injured cannot be compared in different Member States. In addition, the 

definition of injury severity differs among member states, so that an accident or 

casualty which would be recorded in one country might not be recorded in 

another, while an accident or casualty which might be recorded as ‘serious’ in 

one country might be recorded as ‘slight’ in another, 

The result is that at present international comparisons of the level of road safety 

rely almost exclusively upon the analysis of data for fatal accidents and 

fatalities. This is unsatisfactory, for by most criteria non-fatal accidents and 

casualties impose a burden on society that is at least as great as fatal accidents 

and fatalities. For example, using the British Government’s cost-benefit value of 

prevention of road accidents, in 2005 fatal road accidents accounted for 37% of 

the cost-benefit value of preventing accidents that involved personal injury, and 

this falls to 27% when accidents involving material damage only are included. 

Thus, while fatal accidents and casualties form a major part of the burden of 

road accidents in a country, they by no means represent the totality. 

In order to overcome the inconsistencies of reporting and permit non-fatal 

accidents and casualties to be analysed meaningfully, this Task has attempted 

to: 

1. estimate the under-reporting level for each casualty severity (seriously 

injured, slightly injured) by developing a uniform methodology and applying it 

in eight EU countries. 

2. estimate in each country the number of casualties according to a new 

common measurement unit. 

The results from this Task will expand the scope of CARE-based road accident 

analyses. By allowing consideration of road safety to extend beyond its current 

focus on fatal accidents, the increased size of the data sets available for 

analysis will reduce the effects of chance, thereby permitting more detailed 

analyses to be carried out. 

The structure of the work of Task 1.5 has been as follows: 

1.5.1 Development of common methodology 

1.5.2 Execution of National studies on under-reporting; these will be 

conducted by WP1 partners in eight EU countries (Austria, Czech 

Republic, France, Greece, Hungary, Netherlands, Spain, United 

Kingdom) 




1.5.3 Elaboration of National under-reporting coefficients 

1.5.4 Adoption of the common definition of hospitalised persons 

Work has now been completed on all subtasks and this report presents the full 

results. Details of Subtasks 1.5.1 and 1.5.2 were presented in the Interim 

Progress Report (Deliverable D1.6), and these are summarised in the 

remainder of this section for completeness. The common methodology specified 

that the final data from the National studies should be provided in a certain 

common format, and Section 3 presents results prepared using a common 

format for all studies. These results allow Subtasks 1.5.3 and 1.5.4 to be 

completed in Section 4. Section 5 discusses the conclusions that can be drawn 

from this research and presents recommendations for future work in this area. 

Each of the national studies represents a substantial body of work, with details 

varying between countries because of variations in the type of data that are 

available. Appendix A presents reports from the national studies, including 

results of more detailed analyses carried out in individual countries. 

The national studies have been carried out for a specific purpose within the 

SafetyNet project, but it should be recognised that the linked data sets have a 

potential value that extends well beyond this purpose. In particular, clinical 

details from the original medical records are now available for large numbers of 

casualties reported by the police, and this enhanced information has 

considerable research potential – and indeed would be very difficult and 

expensive to collect in a specific research project. These potential applications 

will not be discussed farther in this report, but should be borne in mind when 

assessing the costs and benefits of this type of study. 

Terminology 

The CARE database uses three categories of injury severity: fatal, serious and 

slight. As it combines various national databases, the precise meaning of these 

terms varies by country, and indeed this is one of the main reasons for carrying 

out this research. Further, many studies have found that police officers 

sometimes do not apply the definition current in their country correctly when 

reporting an accident. Arguably, one might use quotation marks as a reminder 

of the possibility of mis-reporting in CARE, for example using “serious” to refer 

to casualties recorded in CARE as being seriously injured but including some 

who were actually fatally or slightly injured. 

The possibility of mis-reporting of injury severity in police accident data is widely 

recognised. Most readers of this report will be aware of this issue so it seems 

unnecessary to use quotation marks. 

It is useful at this point to mention one adaptation of the standard terminology 

that will be made in Section 4.5. A criterion is developed in Section 4 for 

defining a serious casualty in a uniform way in different countries, and this 

category will be termed serious* to differentiate it. 




2.1 Background 

As explained above, road accident reporting systems and standards differ 

among Member States, so it has been necessary to identify a common 

international standard to provide a benchmark with which to compare accident 

data from each country. The benchmark is achieved by developing the method 

that has been used several times in several countries to study the level of 

under-reporting. This consists of comparing: 

° those road accident victims who have been recorded by the police in the 

national accident database, with 

° those who have been recorded in medical records maintained by hospitals. 

Fortunately, medical authorities are farther advanced than transport authorities 

with establishing international recording systems, in particular the International 

Classification of Diseases (ICD) and Abbreviated Injury Scale (AIS) coding 

systems. Hence, the basic approach adopted for this Task consists of 

° linking road accident and medical databases, to investigate the level of 

under-reporting and, equally importantly, to copy details from the medical 

record of each linked casualty to the corresponding record in the accident 

database (Task 1.5.2) 

° comparing the distributions of linked and unlinked casualties from the 

national studies by Maximum AIS (MAIS) and length of stay in hospital (Task 

1.5.3) 

° defining a new injury classification based on the most appropriate medical 

variable(s) and calculating national coefficients to estimate ‘true’ casualty 

totals from the numbers recorded in CARE (Task 1.5.4) 

This has been an ambitious programme of work, and perhaps the greatest 

potential obstacle that was faced has been the difficulty of achieving access to 

suitable databases of medical information. These databases should ideally 

cover distinct, geographically well-defined regions, so that one may be confident 

that any road accident casualty recorded in the medical databases should – in 

the absence of under-reporting - also be recorded in the accident database for 

that region. There may also be ethical problems, since the databases may 

contain details that allow individuals to be identified. On the other hand, the data 

needed for the study are anonymous, the only personal details being age and 

sex which are needed to link to the accident records. 

The linkage of road accident and medical databases is a relatively demanding 

task, involving a degree of judgement when specifying the differences that may 

be tolerated when deciding whether a pair of records actually refer to the same 

casualty. The level of prior experience of this type of work varied widely among 

the partners. 

Similar studies had already been carried out in a number of other countries, 

such as the series of studies that was carried out in Western Australia (e.g. 




Lopez et al, 1999). Indeed, in the USA this type of linkage has been carried out 

routinely for over a decade as part of the Crash Outcome Data Evaluation 

System (CODES) on State-wide data sets from 29 States (see http://wwwnrd.nhtsa.dot.gov/departments/nrd-30/ncsa/CODES.html). 

Thus, this type of 

linkage has been applied in a wide range of contexts, although the technical 

details can vary between studies. The SafetyNet application is thought to be the 

first time that the technique has been used to develop a common benchmark for 

comparing accident data from different countries. 

Road accident victims with only slight injuries may not require significant 

medical treatment and hence would not be recorded in any medical database. 

The common methodology does not cover such cases, as it uses only official 

records. In the longer term it may be possible to develop survey-based methods 

to include them. 

Ideally, under-reporting studies would be carried out in all countries that supply 

data to the CARE database, so the fact that the national studies will only involve 

8 countries also presents difficulties. The pragmatic solution originally 

envisaged had been to use information provided by Governmental Experts 

(including the existence of corresponding data or studies in their countries) to 

draw up a system of analogues, i.e. 

to identify, for each country without a national study, which of the 

countries with a national study it most closely resembles in terms of its 

national accident reporting system, 

to generalise the coefficients estimated for each country with a national 

study to all analogous countries. 

The information provided by Governmental Experts has not been sufficient for 

this to be done. Section 4 will review the data collected by the national studies 

and consider whether an alternative solution may exist. 

Even if such a solution could be identified, any approach which generalises from 

studies in a minority of EU Member States is not fully satisfactory, and the longterm 

aim should be to conduct comparable studies in all Member States. 

The common methodology 

The broad problem that is addressed by Task 1.5 was described in the previous 

section, which also gave a general account of the approach that has been 

adopted. The common methodology that was finalised in May 2005 will now be 

presented. Problems experienced when applying the methodology are recorded 

in the reports in the Appendix and summarised in Section 4. 

There are broadly two alternative methods for collecting the information from 

hospitals to be compared with accident data recorded by the local police: 

1. Regional or national medical authorities may routinely assemble databases 

of records from which suitably detailed medical data can be extracted, 




2. Medical data can be collected specifically for this project. 

In order to carry out option 2, a representative sample of hospitals that receive 

accident victims would be selected, and the approvals needed for the data 

collection obtained from the medical authorities. During the period of data 

collection, project staff would regularly visit the hospital departments that 

receive accident victims. They would sift the records held in these departments 

to identify those people whose presenting history indicates that they had been 

injured in a road accident. A range of details would be recorded for each of 

these people, and subsequently entered in the project database. 

The number of hospitals and the length of the data collection period would 

depend upon the funding available. In view of the limited budget allocated to 

Task 1.5, the collection of data in the volumes required to derive statistically 

reliable results may well not be affordable. In this case, only option 1 would be 

feasible. 

With either option, the medical records are cross-checked regularly with the 

police accident records. The checking takes account of the catchment 1 area of 

each hospital, comparing the hospital records with police accident records only 

for that area. The aim is to identify all cases where the same person is present 

in both sets of records. Personal names are unlikely to be available in both data 

sets, in which case this process must be based on factors common to both data 

sets such as the casualty’s age, sex and mode of travel, together with accident 

circumstances such as date, time and location. 

The outcome is a combined set of police and medical data in which these 

matched cases are marked. The matching process should make allowance for 

minor errors in the recording of personal details, for example small 

discrepancies in age between the two sources. The process used may need to 

vary in detail from country to country to allow for local data and facilities. 

Once the cross-checking of medical and police records has been completed, 

the proportion of accident casualties that has been reported by the police can 

be calculated. This will provide the level of under-reporting of casualties in the 

police data. This is likely to vary with type of accident (e.g. in relation to the 

number and type of vehicles involved), which should also be examined. 

The medical data collected under either option must include those details that 

are needed to cross-check with police records. The principal extra item of data 

is whether or not the casualty was admitted to hospital, and if so for how long. 

Medical details should also be collected, provided ethical approval has been 

given by the medical authorities. The aim will be to record the maximum AIS 2 

score for each body region. Under option 2 it will be possible to do this directly 

1 The area around the hospital from which accident victims are normally brought to the hospital 

for initial treatment 

2 Abbreviated Injury Scale, ranging from 1 for minor injuries to 6 for injuries that are currently 

untreatable 




from the hospital records if the data collection staff have sufficient expertise, or 

it may be necessary to transcribe sufficient details for more expert staff to 

assess the AIS scores subsequently. 

The combined police and medical data set are used to compile two 3- 

dimensional matrices of casualty counts. Matrix 1 is based on a casualty’s 

length of stay in hospital, Matrix 2 is based on the severity of their injuries as 

summarised by the MAIS score (the maximum of the AIS scores per body 

region, as described in the following section). Road user type is identified from 

police data, as it is often poorly recorded in medical records. 

Matrix 1 

road user type Length of Stay police coding 

car occupant 

pedestrian 

pedal cyclist 

motorcyclist 

other 

X 

out-patient 

overnight 

1-3 days 

>3 days 

not coded (not 

matched in medical 

records) 

X 

fatal 

serious 

slight 

not coded (not matched 

in police records) 

Matrix 2 

road user type MAIS police coding 

car occupant 

pedestrian 

pedal cyclist 

motorcyclist 

other 

X 

1-6 

not coded (not 

matched in medical 

records) 

X 

fatal 

serious 

slight 

not coded (not matched in 

police records) 

The ‘Common Methodology’ defines the strategy to be followed by each 

national study and the numerical outputs, but not the details. It was recognised 

at the outset that the type and availability of data varies from country to country, 

so it would be impossible to be more prescriptive. Inevitably, a range of linkage 

methods were applied, and Section 4 considers the implications. Full details of 

the individual studies and their linkage methods are provided in the Appendix. 

The definition of Length of Stay had to be tightened in the course of the national 

studies: 

Out-patient (0 nights, 1 part-day) 

Overnight (1 night, 2 part-days) 

1-3 days (2-4 nights) 

>3 days (>4 nights) 

The Greek study could not apply this definition exactly because of limitations in 

the source data, as explained in Section 7.4. 




2.3 Injury and severity coding 

It is useful at this point to review the way in which details of injuries are coded, 

as this is the basis for summarising the overall severity of a casualty’s injuries 

via the MAIS. The details have evolved over many years, but there is a wellestablished 

international coding system. 

The Abbreviated Injury Scale (AIS) is a specialised trauma classification of 

injuries based mainly on anatomical descriptors of the tissue damage caused by 

the injury (EGISM, 2004). The AIS classification system was designed to 

distinguish between types of trauma of clinical importance as well as types of 

trauma of interest to vehicle designers and research engineers. It has been 

shown to provide a good basis for valid measurement of probability of death. 

The AIS has two components: (1) the injury descriptor (often referred to as the 

‘pre-dot’ code) which is a unique numerical identifier for each injury description; 

and (2) the severity score (can be referred to as the ‘post-dot’ code). The 

severity score ranges from 1 (minor) to 6 (currently untreatable), and is 

assigned to each injury descriptor. The AIS is based on anatomical injury, and 

not on physiological parameters (of the person injured). It implies that there is 

only a single AIS severity score for each injury, for any one person. The AIS 

scores injuries and not the consequences of injuries; it is not a measure of 

impairments that result from the injury. The AIS severity code is not simply a 

ranking of expected mortality from injury; it is based on potential for mortality but 

also on the diagnostic certainty, rapidity, duration, complexity and expected 

effectiveness of resolution with or without existing therapy (AAAM, 1990). The 

MAIS is the maximum AIS of all injury diagnoses for a person 

The MAIS can be estimated directly by trained staff, but alternatively it can be 

derived from other classifications. The International Classification of Diseases 

(ICD) is a system designed to promote international comparability in the 

collection, processing, classification, and presentation of mortality statistics 

(DHHS; NCHS, 2007; WHO, 1992). It provides a way to classify medical terms 

reported by physicians, medical examiners and coroners on death certificates, 

also data from physicians’ offices and hospital inpatient and outpatient records, 

so that they can be grouped together for statistical purposes. 

In the United States, as in many other countries, the ICD is used to code and 

classify mortality data from death certificates. The ICD Clinical Modification 

(CM) is used to code non-fatal-injury data from medical records (i.e., hospital 

records, emergency department records, and physician office data). From 1979 

to 1998, injury-related fatalities were coded using the 9th revision of the ICD 

(ICD-9) external cause of injury codes, more commonly referred to as E codes. 

In 1999, the 10th revision of the ICD (ICD-10) was implemented for coding 

deaths (DHHS). There is as yet no Clinical Modification of ICD-10. Countries 

have decided individually if and when to migrate from ICD-9 to ICD-10. 

Under ICD-9, the external cause of injury or death was assigned an E code 

ranging from E800.0 to E999.9 based on information documented on the death 




certificate. External cause of injury codes describe the circumstances, such as a 

motor vehicle crash, drowning, or suffocation, as well as the intent of the injury 

(i.e. unintentional, homicide, suicide, intent undetermined, or other). In 1999, 

along with ICD-10, the injury E codes for fatalities were replaced with V, W, X, 

and Y cause codes along with special U codes for terrorism. 

The ICD is developed collaboratively between the World Health Organization 

(WHO) and 10 international centres, to ensure that medical terms reported on 

death certificates are internationally comparable and lend themselves to 

statistical analysis. The ICD has been revised approximately every 10 years 

since 1900. These revisions reflect advances in the medical field and changes 

in our understanding of disease mechanisms and terminology, and are 

designed to maximise the amount of information and flexibility a code can 

provide. ICD-10 more closely reflects current medical knowledge than ICD-9. 

One approach for using the ICD for severity assessment has been to develop 

the ICDMAP software that translates ICD-9-CM coded discharge diagnoses into 

AIS pre-dot codes, injury descriptors and severity scores (MacKenzie et al. 

1997). The mapping does result in some loss of information due to differences 

in the injury classification systems. Resulting severity scores referred to as 

ICD/AIS scores are considered to be conservative measures of injury severity. 

Until recently there was no software to convert the current ICD-10 codes to AIS. 

In 2006, Dr. Maria Seguí-Gómez from the Universidad de Navarra (Spain) 

developed a program that allows AIS to be coded from ICD-10 (ECIP, 2006). 

There is as yet no validation study that compares the effect on severity trends of 

changing from ICD-9-CM to ICD-10. An empirical adjustment is derived in 

Section 4.4 for road accident casualties with MAIS>=3. 

Another approach is to estimate the severity of an injury based on the estimated 

probability of surviving that injury, P(survival), known as ICISS (Osler et al. 

1996). To calculate this, first survival risk ratios (SRRs) are calculated for each 

diagnosis code as the proportion of cases with that diagnosis code who did not 

die. SRRs are calculated by dividing the number of survivors among patients 

with a specific ICD by the total number of patients with that ICD code. Each 

case is then assigned an ICISS, which is the product of SRRs of all their 

diagnoses. The resulting ICISS are estimates of P(survival) which range from 0 

(unsurvivable) to 1 (certain to survive) (Stephenson et al, 2005). 

The ICISS has also problems that need to be addressed, such as the fact that it 

is in some measure database-specific, or depends upon other injuries in 

multiple trauma cases. Nonetheless, the development of this approach to injury 

severity assessment is on-going and shows great promise. It could provide a 

valid alternative to MAIS in future. 




3. Results 

Two matrices are prepared in each national study for the purpose described in 

the Introduction: to estimate Conversion Factors that can be applied to police 

accident statistics, e.g. as held in the CARE database, in order to estimate 

national casualty totals according to two criteria, Length of Stay or MAIS. The 

calculations are set out in detail in Section 3.1 using data from the UK national 

study. Certain assumptions are required for the calculations, and these are 

presented and discussed in the context of these data. 

Results from the other studies are presented in less detail in Section 3.2. More 

detailed results of interest from individual studies are presented in the Appendix, 

as part of the technical descriptions of the studies. 

First, a general overview of the national studies is given in Table 1, while Figure 

1 illustrates the eight countries involved. 

Table 1: Summary details of studies 

Country Study area Period Coding of MAIS 

Austria National 2001 From ICD10 

Czech Republic Local (Kromeriz, central 2003 - 2005 From ICD10 

Moravia) 

France Regional (Département 1996 - 2003 Coded directly 

of the Rhône) 

Greece Regional (Corfu) 1996 - 2003 From ICD9 

Hungary Local (part of Budapest) Aug 2004 - Jan 2006 Coded directly 

Netherlands National 1997 - 2003 From ICD9 

Spain 

Regional (Castilla y July - Dec 2005 From ICD9 

Leon) 

United Kingdom Regional (Scotland) 1997 - 2005 From ICD10 

The fundamental assumption 

The fundamental assumption that underlies the calculations in the following 

Sections is that the medical and police data have been linked correctly, i.e. the 

links that have been made are valid, whereas records that have not been linked 

genuinely refer to different people. Clearly, the validity of this assumption 

depends upon the accuracy of the data in the two sets of records that are used 

for the linking process, but it is inescapable. This assumption needs to be borne 

in mind when reading the explanation of the calculations. Further assumptions 

are introduced as appropriate. 

The accuracy of the linkage achieved could only be checked rigorously with 

access to the personal identifiers in the two sources of information for at least a 

subset of records. Such highly confidential information was not available to any 

of the national studies. 




Figure 1: Countries where studies were carried out 

3.1 Results from the UK study 

The UK study linked records from the Scottish Hospital Inpatient System for 

1997-2005 with STATS19 police accident records from Scotland. Full details are 

provided in Section 7.8. The results can be summarised as follows: 

Police 

Hospital 26,198 casualties in 

SHIPS and STATS19 

Not 

hospital 

151,165 casualties in 

STATS19 but not in 

SHIPS 

Not police 

20,672 casualties in 

SHIPS but not in 

STATS19 

unknown number of 

casualties neither in 

SHIPS nor in STATS19 

The results of the linkage are summarised by road user type and policereported 

casualty severity in Table 2. 




Table 2: Linkage results, by road user type 

Police 

Not Grand 

Road user Fatal Serious Slight police Total 

Hospital Car occupant 186 8,776 5,280 7,238 21,480 

Motorcyclist 26 1,979 620 2,736 5,361 

Pedal cyclist 18 697 422 4,018 5,155 

Pedestrian 180 5,255 2,008 2,888 10,331 

Unknown 0 0 0 1,229 1,229 

Other 17 727 434 2,563 3,741 

Not 

hospital 

Length of Stay 

Subtotal 427 17,434 8,764 20,672 47,297 

Car occupant 1,471 6,891 91,578 99,940 

Motorcyclist 334 1,393 5,225 6,952 

Pedal cyclist 79 709 6,480 7,268 

Pedestrian 516 2,578 20,822 23,916 

Other 144 1,260 14,229 15,633 

Subtotal 2,544 12,831 138,334 153,709 

Grand Total 2,971 30,265 147,098 20,672 201,006 

The Length of Stay data will be analysed first. The overall results of the linkage 

are shown in Table 3, omitting fatal casualties. The proportion of casualties who 

were not reported by the police is lower among the more severely injured. The 

Length of Stay is reported for all SHIPS cases, so there are no unknown cases. 

Table 3: The linkage results, by Length of Stay 

Length of Stay Police 

Not police % not reported 

Serious Slight 

by police 

Hospital Outpatient 1,152 1,179 3,596 61% 

Overnight 4,434 4,336 7,219 45% 

1-3 days 4,690 2,132 4,880 42% 

>3 days 7,158 1,117 4,977 38% 

Not 

12,831 138,334 

hospital 

Grand Total 30,265 147,098 20,672 

The ‘police, not hospital’ casualties did not attend hospital, assuming that the 

record-linkage is perfect, so they are shown in Table 4 as ‘not in hospital’. The 

‘hospital, not police’ casualties must be divided between the serious and slight 

categories so as to simulate the severity coding that the police would have used 

if they had been aware of these accidents, rather than the actual coding of 

serious. This is achieved by distributing the casualties for each Length of Stay 

(LoS) pro rata between the serious and slight categories. The calculation is as 

follows: 




Let: 

ser(i) = number of serious casualties reported by police with LoS=i 

sli(i) = number of slight casualties reported by police with LoS=i 

hnp(i) = number of casualties in hospital with LoS=i but not reported by police 

then estimated total of serious(i) = ser(i)*[ser(i)+sli(i)+hnp(i)] / [ser(i)+sli(i)] 

estimated total of slight (i) = sli(i)*[ser(i)+sli(i)+hnp(i)] / [ser(i)+sli(i)] 

Table 4 presents the results. Capture/recapture methods have been used in 

some studies to estimate the number in the blank cell, i.e. casualties that were 

neither reported by police nor hospital. This has not been attempted here, but 

once a method was agreed it would be simple to update the calculation. 

Table 4: Estimated results, by Length of Stay 

Length of Stay Police Not police Estimated total 



Outpatient 1,152 1,179 3,596 2,929 2,998 

Overnight 4,434 4,336 7,219 8,084 7,905 

1-3 days 4,690 2,132 4,880 8,045 3,657 

>3 days 7,158 1,117 4,977 11,463 1,789 

Not in hospital 12,831 138,334 12,831 138,334 

Total 30,265 147,098 20,672 43,352 154,683 

The results show that, corresponding to each STATS19 serious casualty, 

11,463/30,265=0.38 casualties were in hospital for more than 3 days, and 

corresponding to each slight casualty another 1,789/147,098=0.012 casualties 

were in hospital for more than 3 days. Such conversion factors can be used to 

estimate casualty totals from STATS19 casualty totals. For example, if serious 

casualties were to be defined as those staying more than 3 days in hospital then 

the actual total could be estimated as: 

N = 0.38 x number of serious casualties reported by the police + 

0.012 x number of slight casualties reported by the police 

Note, however, that the calculation of the factors depends upon the period of 

the casualty data, as changes over time in hospital procedures are likely to 

affect the Length of Stay for any particular casualty. The conversion factors 

calculated by road user type are presented in Table 5 and illustrated in Figure 2. 

Table 5 is the first of the standard tables that will be used to compare results 

from the national studies. 




Table 5: Conversion Factors based on Length of Stay 

Length of Car Occupant Motorcyclist Pedal Cyclist Pedestrian Other All 

Stay Serious Slight Serious Slight Serious Slight Serious Slight Serious Slight Serious Slight 

Outpatient/ 

overnight 

0.33 0.06 0.34 0.14 1.16 0.25 0.28 0.08 0.42 0.06 0.36 0.074 

1-3 days 0.21 0.02 0.36 0.08 0.68 0.06 0.26 0.03 0.29 0.02 0.27 0.025 

>3 days 0.32 0.01 0.49 0.03 0.40 0.01 0.41 0.02 0.54 0.01 0.38 0.012 

All 0.86 0.09 1.20 0.25 2.24 0.33 0.95 0.13 1.25 0.10 1.01 0.111 

>=1 day 0.53 0.03 0.85 0.11 1.08 0.08 0.66 0.05 0.82 0.03 0.64 0.037 

Figure 2: Conversion Factors based on Length of Stay 

1.0 

>3 days (serious) 1-3 days (serious) 3 days (slight) 1-3 days (slight)


Table 6: The linkage results, by MAIS 

Police 

Not police % not reported 

MAIS Serious Slight 

by police 

Hospital 1 3,823 3,642 6,294 46% 

2 8,336 2,473 8,050 43% 

3 3,139 412 2,227 39% 

4 226 33 220 46% 

5 75 1 44 37% 

6 197 19 134 38% 

9 1,638 2,184 3,703 49% 

1-9 17,434 8,764 20,672 

Not 

hospital 12,831 138,334 

The MAIS scores have been assigned from the ICD10 injury codes for each 

case. MAIS 9 is a code generated by the mapping algorithm that represents not 

known, i.e. the ICD10 codes were not sufficiently detailed to assign an MAIS 

score, e.g. ‘head injury’. 9.4% of serious casualties have MAIS 9, and 24.9% of 

slight casualties. These casualties appear on the whole to have relatively minor 

injuries, for example the incidence of MAIS 9 is lower among serious casualties 

than among slight and the % not reported by police is greater than for MAIS 1. 

The discussion in Section 2.3 suggests that the lack of a Clinical Modification of 

ICD-10 may explain the incidence of MAIS 9 casualties. 

Excluding these cases would introduce one type of bias, tending to raise the 

apparent reporting level, while treating them all as MAIS 1 would introduce 

another type as they may well include cases with an actual MAIS of at least 3. 

On the whole, it appears preferable to treat the MAIS 9 cases as MAIS 1, so in 

the remainder of this report MAIS 9 has been grouped with MAIS 1. 

The estimation process is more complex for MAIS than for Length of Stay in 

another respect. It was known that ‘police, not hospital’ cases had zero Length 

of Stay (assuming the record-linkage is perfect), but the MAIS of these cases 

must be estimated. These casualties have not required in-patient treatment, so 

it seems unlikely that MAIS will have exceeded 3. On the other hand, some 

MAIS 2 casualties could well be treated as outpatients, or in local doctors’ 

surgeries. It is reasonable to assume that all of these casualties had MAIS 1 or 

2, but that they cannot be distributed reliably between 1 and 2. As with the 

calculation for Length of Stay, casualties that were not reported by the police 

are then distributed pro rata at each MAIS level to simulate the police severity 

coding, with the results shown in Table 7. Note that the calculation is carried for 

each MAIS value separately, then the results for values of 1, 2 and 9 are 

summed to form the first row. As noted previously, the SHIPS data cannot 

estimate reliably the actual number of slight casualties with MAIS 1 or 2. 




Table 7: Estimated results, by MAIS 

Estimated distribution 

of police, 

not hospital 

Estimated total 

Not 

reported 

by police Serious Slight Serious Slight 

Reported by 

police 

MAIS Serious Slight 

1 or 2 13,797 8,299 18,047 12,831 138,334 37,647 1 153,661 1 

3 3,139 412 2,227 0 0 5,108 670 

4 226 33 220 0 0 418 61 

5 75 1 44 0 0 118 2 

6 197 19 134 0 0 319 31 

All 17,434 8,764 20,672 43,610 1 154,425 1 

1 likely to be underestimated 

Thus, it is estimated that for each serious casualty in the STATS19 records 

there are actually 5,963/30,265=0.20 casualties with MAIS>=3. Further, a small 

proportion of slight casualties in the SHIPS records actually had MAIS>=3, 

764/147,098=0.005 per slight casualty. If serious casualties were to be defined 

as those with MAIS>=3 then the actual total (all types of road user) could be 

estimated as: 

N 1 = 0.20 x number of serious casualties reported by the police + 


Again, these values are averages for the 1997-2005 period, and slightly 

different values would apply for, say, the 2003-05, period. 

Table 8 presents the conversion factors by road user type. The results are 

illustrated in Figure 3, grouping MAIS>=3 together (factors for slight casualties 

are multiplied by 25 to facilitate visual comparison). The Figure emphasises that 

car occupants and pedestrians are recorded more fully in the Scottish STATS19 

data than pedal cyclists. This is the other standard table that will be used to 

compare results from the national studies. Note, however, that certain of these 

results will need to be adjusted for reasons that are explained in Section 4.3. 

Table 8: Conversion Factors based on MAIS 

Car Occupant Motorcyclist Pedal Cyclist Pedestrian Other All 

MAIS Serious Slight Serious Slight Serious Slight Serious Slight Serious Slight Serious Slight 

1 or 2 1 1.15 1.03 1.34 1.13 2.54 1.24 1.05 1.03 1.62 1.06 1.24 1.04 

3 0.13 0.00 0.25 0.01 0.26 0.01 0.18 0.01 0.23 0.01 0.17 0.00 

4 0.01 0.00 0.01 0.00 0.02 0.00 0.03 0.00 0.01 0.00 0.01 0.00 

5 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 

6 0.01 0.00 0.01 0.00 0.01 0.00 0.02 0.00 0.01 0.00 0.01 0.00 

All 1 1.30 1.03 1.61 1.14 2.83 1.25 1.28 1.04 1.88 1.07 1.44 1.05 

>=3 0.15 0.00 0.27 0.01 0.29 0.01 0.23 0.01 0.26 0.01 0.20 0.01 

1 factors are likely to be underestimated 




pedal cyclist 

Figure 3: Conversion Factors based on MAIS 

other 

motorcyclist 

car occupant 

pedestrian 

1 or 2 (serious) 

>=3 (serious) 

>=3 (slight)*25 

All 

0.0 0.5 1.0 1.5 2.0 2.5 

3.2 Results from the other studies 

This Section brings together results from the other national studies. The 

methods of calculation are identical to those used in the UK study, so the results 

are presented in less detail. The same structure is used in each case, but 

variations in the precise content of the data received mean that there are 

differences in detail. 

The overall results from the various national studies are compared in Section 4. 

Austrian study 

The Austrian study was carried out with national data for the road accidents that 

occurred in the year 2001. The medical information comes from the national 

hospital discharge database, which combines administrative and medical data 

for all in-patients in the 270 Austrian hospitals. No out-patients are recorded in 

this database. 

The medical records contain no information about the road user type of a 

casualty, so it is only possible to calculate overall conversion factors, i.e. not by 

road user type. Moreover, the records contain only one ICD code per casualty, 

whereas the software used to estimate MAIS uses up to 3 ICD codes. 

Consequently it has only been possible to calculate conversion factors by 

Length of Stay. 




Police 

Fatally 

Injured 

Table 9: The linkage results, Austria 

Severity 

unknown 

Seriously 

Injured 

Slightly 

Injured 

Not 

police 

Grand 

Total 

Hospital 104 1,382 2,882 1,689 12,010 18,067 

Not hospital 854 5,187 5,325 39,800 51,166 

Grand Total 958 6,569 8,207 41,489 12,010 69,233 

Table 10: Conversion Factors based on Length of Stay, Austria 

Czech study 

Length of 

Stay 

Severity 

unknown 

Seriously 

Injured 

Slightly 

Injured 

Overnight 0.20 0.17 0.051 

1 - 3 days 0.26 0.33 0.051 

>3 days 0.18 0.53 0.025 

All 0.64 1.03 0.127 

>=1 day 0.45 0.86 0.076 

The Czech study was carried out for the district of Kromeriz in the years 2003 – 

2005. The town lies in central Moravia, about 70 km from Brno, and has a 

population of 30,000 inhabitants. There is one hospital, which is the source of 

the medical data. 

Table 11: The linkage results, by road user type, Czech Republic 

Police 

Seriously 

injured 

Not 

police 

Grand 

Total 

Slightly 

injured 

Hospital Car occupant 18 126 76 220 

Motorcyclist 4 14 15 33 

Pedal cyclist 10 49 422 481 

Pedestrian 8 18 62 88 

Other 1 10 0 11 

Not 

hospital 

Subtotal 41 217 575 833 

Car occupant 64 441 505 

Motorcyclist 18 54 72 

Pedal cyclist 20 110 130 

Pedestrian 20 44 64 

Other 7 38 45 

Subtotal 129 687 816 

Grand Total 170 904 575 1,649 




Table 12: Conversion Factors based on Length of Stay, Czech Republic 

Length of Car occupant Motorcyclist Pedal cyclist Pedestrian Other All 


Overnight 0.10 0.28 0.19 0.34 0.71 2.57 0.19 0.98 0.00 0.19 0.23 0.72 

1-3 days 0.09 0.06 0.00 0.04 0.30 0.15 0.07 0.16 0.13 0.02 0.11 0.08 

>3 days 0.11 0.01 0.09 0.01 0.50 0.02 0.31 0.02 0.00 0.00 0.19 0.02 

All 0.29 0.35 0.28 0.39 1.51 2.74 0.57 1.16 0.13 0.21 0.53 0.82 

>=1 day 0.19 0.07 0.09 0.06 0.80 0.17 0.38 0.18 0.13 0.02 0.30 0.09 

Table 13: Conversion Factors based on MAIS, Czech Republic 

MAIS Car occupant Motorcyclist Pedal cyclist Pedestrian Other All 

Serious Slight Serious Slight Serious Slight Serious Slight Serious Slight Serious Slight 

1 or 2 1 0.97 1.11 1.03 1.17 1.11 3.50 1.05 1.77 0.88 1.00 1.07 1.56 

3 0.07 0.01 0.05 0.01 0.30 0.04 0.31 0.04 0.13 0.00 0.15 0.02 

4 0.01 0.00 0.00 0.00 0.17 0.00 0.00 0.00 0.00 0.00 0.03 0.00 

5 0.02 0.00 0.05 0.00 0.03 0.00 0.04 0.00 0.00 0.00 0.03 0.00 

6 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 

All 1 1.08 1.12 1.12 1.18 1.61 3.54 1.40 1.80 1.00 1.00 1.28 1.58 

>=3 0.11 0.01 0.09 0.01 0.50 0.04 0.35 0.04 0.13 0.00 0.21 0.02 


French study 

The French study was carried out with data for the eight years 1996-2003 from 

the département of the Rhône, an area of 1.6 million inhabitants consisting of 

the city of Lyon, its suburbs and a rural area to the north. A road trauma registry 

has operated in the département since 1995, covering all road accident 

casualties who seek medical attention in health facilities. These data were 

linked with police records, and the results of the linkage are summarised by 

road user type in Table 14. 




Table 14: The linkage results, by road user type, France 

Police 

Seriously Slightly 

Not 

police 

Grand 

Total 

Road user injured injured 

Hospital Car occupant 1,703 11,195 28,212 41,110 

Motorcyclist 983 2,814 11,702 15,499 

Pedal cyclist 153 591 9,982 10,726 

Pedestrian 734 2,334 4,333 7,401 

Other 74 682 2,250 3,006 

Not 

hospital 

Subtotal 3,647 17,616 56,479 77,742 

Car occupant 544 7,509 8,053 

Motorcyclist 304 1,718 2,022 

Pedal cyclist 59 318 377 

Pedestrian 279 1,430 1,709 

Other 29 525 554 

Subtotal 1,215 11,500 12,715 

Grand Total 4,862 29,116 56,479 90,457 

Table 15: Conversion Factors based on Length of Stay, France 

Length of Car occupant Motorcyclist Pedal cyclist Pedestrian Other All 


Overnight 1.12 1.91 1.28 2.61 4.82 9.82 0.79 1.44 1.16 2.18 1.21 2.23 

1-3 days 0.18 0.04 0.19 0.11 0.58 0.31 0.14 0.07 0.41 0.05 0.20 0.06 

>3 days 0.48 0.03 0.67 0.09 1.22 0.13 0.48 0.08 0.90 0.05 0.57 0.05 

All 1.78 1.98 2.14 2.81 6.62 10.26 1.41 1.59 2.47 2.28 1.98 2.34 

>=1 day 0.66 0.07 0.86 0.20 1.80 0.44 0.62 0.14 1.31 0.10 0.77 0.11 

Table 16: Conversion Factors based on MAIS, France 

Car occupant Motorcyclist Pedal cyclist Pedestrian Other All 


1 or 2 1 1.32 2.38 1.35 3.13 4.69 10.39 1.01 1.90 1.52 2.67 1.43 2.69 

3 0.35 0.03 0.69 0.11 1.64 0.27 0.43 0.08 0.69 0.05 0.52 0.05 

4 0.12 0.00 0.10 0.01 0.26 0.00 0.10 0.01 0.30 0.01 0.12 0.01 

5 0.05 0.00 0.05 0.00 0.07 0.00 0.03 0.00 0.07 0.00 0.05 0.00 

6 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 

All 1 1.84 2.41 2.18 3.25 6.67 10.66 1.58 2.00 2.58 2.73 2.11 2.75 

≥3 0.51 0.03 0.83 0.12 1.97 0.27 0.57 0.10 1.06 0.06 0.68 0.06 





Greek Study 

The Greek study was carried out with data for 1996–2003 from the Island of 

Corfu. The island has a population of approximately 110,000 people and is 

located in the North Ionian Sea. The study was based on data from the Greek 

Emergency Department Injury Surveillance System (EDISS) which collects data 

from the Regional Hospital of Corfu (the only public hospital on the island) and 

the Road Traffic Police database. 

Table 17: The linkage results, by road user type, Greece 

Police 

Not Grand 

Road user Seriously injured Slightly injured police Total 

Hospital Bicyclist 2 4 98 104 

Driver 141 615 5,949 6,705 

Passenger 43 191 1,375 1,609 


Unknown 24 118 1,959 2,101 

Not 

hospital 

Subtotal 240 1,022 10,005 11,267 

Driver 54 315 369 

Passenger 26 153 179 


Subtotal 102 546 648 

Grand Total 342 1,568 10,005 11,915 

Table 18: Conversion Factors based on Length of Stay, Greece 

Length of Driver Passenger Pedestrian Unknown All 

Stay Serious Slight Serious Slight Serious Slight Serious Slight Serious Slight 

Overnight 3.83 4.46 2.38 2.97 0.37 2.76 15.65 13.33 3.91 4.63 

1-3 days 1.35 1.22 0.88 0.78 0.85 0.70 0.45 0.94 1.15 1.04 

>3 days 0.55 0.33 0.36 0.20 0.56 0.35 0.10 0.23 0.49 0.30 

All 5.73 6.01 3.62 3.95 1.78 3.81 16.20 14.51 5.55 5.97 

>=1 day 1.90 1.55 1.24 0.98 1.41 1.05 0.56 1.17 1.64 1.34 

Table 19: Conversion Factors based on MAIS, Greece 

Driver Passenger Bicyclist Pedestrian Unknown All 


1 or 2 1 4.62 6.40 3.19 4.42 8.33 20.58 2.49 3.91 11.82 15.17 4.54 6.39 

3 0.47 0.11 0.27 0.08 0.00 1.00 0.31 0.13 0.53 0.07 0.40 0.11 

4 0.06 0.01 0.00 0.00 0.17 0.17 0.13 0.00 0.17 0.02 0.06 0.01 

5 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 

All 1 5.14 6.53 3.46 4.50 8.50 21.75 2.93 4.04 12.51 15.26 5.00 6.51 

>=3 0.53 0.13 0.27 0.09 0.17 1.17 0.45 0.14 0.69 0.09 0.46 0.12 





Hungarian study 

The study in Hungary was carried out with data from the Károlyi Sándor 

Hospital in Budapest, one of the 4 Regional Trauma Centres in the city. The 

accidents occurred between 1 August 2004 and 31 January 2006. 

Table 20: The linkage results, by road user type, Hungary 

Police 

Not police Grand 

Fatally Seriously Slightly 

Total 

Road user injured injured injured 

Hospital Car Occupant 2 67 202 273 544 

Motorcyclist 0 52 44 182 278 

Pedal Cyclist 0 10 7 219 236 

Pedestrian 1 1 45 62 89 197 

Other 0 0 8 31 39 

Subtotal 3 174 323 794 1,294 

Not Vehicle Occupant 54 374 1,359 1,787 

hospital Pedestrian 29 111 238 378 

Subtotal 83 485 1,597 2,165 

Grand Total 86 659 1,920 794 3,459 

1 includes 8 casualties recorded by police as vehicle occupants 

Table 21: Conversion Factors based on Length of Stay, Hungary 

Length of Vehicle occupant Pedestrian All 

Stay Serious Slight Serious Slight Serious Slight 

Overnight 0.12 0.34 0.04 0.17 0.10 0.31 

1-3 days 0.10 0.13 0.08 0.16 0.10 0.13 

>3 days 0.35 0.032 0.37 0.069 0.35 0.038 

Total 0.57 0.50 0.48 0.40 0.55 0.49 

>=1 day 0.45 0.16 0.44 0.22 0.45 0.17 

Table 22: Conversion Factors based on MAIS, Hungary 

Vehicle Occupant Pedestrian All 

MAIS Serious Slight Serious Slight Serious Slight 

1+2 1 0.83 1.28 0.86 1.16 0.84 1.27 

3 0.43 0.04 0.22 0.02 0.38 0.04 

4 0.06 0.00 0.08 0.00 0.06 0.00 

5 0.02 0.00 0.05 0.00 0.03 0.00 

6 0.00 0.00 0.01 0.00 0.00 0.00 

All 1 1.35 1.33 1.21 1.19 1.32 1.31 

>=3 0.52 0.04 0.35 0.03 0.48 0.04 


Note that, as with Table 8, certain of these results will need to be adjusted for 

reasons that are explained in Section 4.3. 




Dutch Study 

The Dutch study was carried out with data for 1997-2003 from the whole of the 

Netherlands. Rather than serious casualties, the data refer to hospitalised 

(Hosp) casualties. 

Table 23: The linkage results, by road user type, the Netherlands 

Police 


Road user Hosp. Slight 

Total 

Hospital Car/van occupant 21,176 4,590 12,570 38,336 

Motorcyclist 3,847 831 3,266 7,944 

Moped 9,230 2,477 11,146 22,853 

Pedal cyclist 10,323 2,732 32,006 45,061 

Pedestrian 3,620 781 3,693 8,094 

Other 539 86 672 1,297 

Not 

hospital 

Subtotal 48,735 11,497 63,353 123,585 

Car/van occupant 1,479 1,309 2,788 


Moped 599 530 1,129 

Pedal cyclist 644 570 1,214 


Other 48 31 79 

Subtotal 3,205 2,826 6,031 

Grand Total 51,940 11,497 66,179 129,616 

Table 24: Conversion Factors based on Length of Stay, the Netherlands 

Length of Car occupant Motorcyclist Moped rider Pedal cyclist Pedestrian Other All 

Stay Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight 

Overnight 0.13 0.017 0.10 0.022 0.14 0.017 0.22 0.029 0.12 0.022 0.08 0.009 0.15 0.021 

1-3 days 0.40 0.036 0.46 0.067 0.49 0.045 0.89 0.083 0.49 0.061 0.39 0.029 0.52 0.051 

>3 days 0.39 0.014 0.68 0.049 0.69 0.030 1.22 0.055 0.71 0.046 0.39 0.014 0.65 0.029 

All 0.92 0.067 1.25 0.138 1.31 0.092 2.33 0.167 1.31 0.129 0.86 0.052 1.32 0.101 

>=1 day 0.78 0.050 1.14 0.116 1.17 0.075 2.11 0.138 1.20 0.107 0.78 0.043 1.17 0.080 




Table 25: Conversion Factors based on MAIS, the Netherlands 

Car occupant Motorcyclist Moped rider Pedal cyclist Pedestrian Other All 

MAIS Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight Hosp. Slight 

1 1 0.19 0.019 0.12 0.021 0.14 0.014 0.23 0.022 0.14 0.020 0.16 0.014 0.18 0.019 

2 0.41 0.029 0.71 0.088 0.72 0.057 1.27 0.099 0.74 0.080 0.42 0.026 0.68 0.056 

3 0.20 0.006 0.31 0.020 0.35 0.014 0.70 0.031 0.33 0.019 0.19 0.007 0.34 0.015 

4 0.02 0.001 0.03 0.001 0.03 0.001 0.05 0.002 0.04 0.002 0.02 0.001 0.03 0.001 

5 or 6 0.02 0.000 0.03 0.001 0.02 0.001 0.02 0.000 0.02 0.000 0.01 0.000 0.02 0.000 

All 1 0.83 0.055 1.20 0.131 1.26 0.086 2.27 0.154 1.26 0.122 0.80 0.048 1.25 0.090 

>=3 0.23 0.007 0.37 0.022 0.40 0.015 0.77 0.033 0.38 0.021 0.23 0.008 0.39 0.016 


Spanish study 

Two linkage studies have been carried out in Spain, one in the city of Barcelona 

and the other in the rural region of Castilla y Leon. It was concluded that the 

results from Barcelona were not representative, so they will not be presented. 

Only overall results could be prepared in either study, i.e. not by road user type. 

The reason is illustrated by Table 26: the road user type is unknown in the 

medical records. The data come from July-December 2005. 

Table 26: The linkage results, by road user type, Castilla y Leon 

Police 

Fatal Serious Slight Not coded 


Total 

Hospital Car occupant 6 176 104 286 



Motor cyclist 2 56 10 68 

Other 1 38 23 62 

Unknown 0 0 0 1,143 1,143 

Not 

hospital 

Subtotal 13 326 154 1,143 1,636 

Car occupant 155 757 3,460 4,372 

Pedestrian 17 105 316 438 


Motor cyclist 20 183 432 635 

Other 37 210 644 891 

Unknown 0 0 2 27 29 

Subtotal 234 1,283 4,933 27 6,477 

Grand Total 247 1,609 5,087 27 1,143 8,113 




Table 27: Conversion Factors based on Length of Stay, Castilla y Leon 


Seriously Slightly 

injured injured 

Overnight 0.01 0.007 

1-3 days 0.19 0.051 

>3 days 0.46 0.050 

All 0.67 0.107 

>=1 day 0.66 0.101 

Table 28: Conversion Factors based on MAIS, Castilla y Leon 


1 or 2 1 1.22 1.06 

3 0.16 0.01 

4 0.08 0.00 

5 0.03 0.00 

6 0.00 0.00 

All 1 1.48 1.07 

>=3 0.26 0.02 





4. Synthesis 

This Section brings together the results of the national studies presented in the 

previous section. First, however, the studies themselves will be reviewed briefly, 

drawing upon the detailed report of each study in the Appendix. 

Predictably, all studies used accident data from national accident databases 

that had been compiled from police accident reports. Most studies used files of 

medical data compiled by national or regional authorities from hospital records. 

The medical files available in the Czech Republic and Hungary were not 

sufficient to carry out the study, however, so the medical data had to be 

assembled from hospital records specifically for these studies. This is clearly a 

more expensive process and has restricted the scale of these studies. The 

national medical file available in Austria only included a subset of the injury 

codes, which hampered that study considerably. 

There were problems of obtaining access to medical records in most studies, 

although these were overcome successfully. It is worth recalling, however, that 

it was originally envisaged that a study would be carried out in Belgium, but this 

had to be aborted when the Belgian partner found it impossible to negotiate 

access to the necessary data. 

Access to anonymised medical records for research purposes is often 

problematical. Modern linkage techniques such as those used in this project, 

however, make these data increasingly valuable. Ways need to be found to 

persuade the custodians of these data to allow them to be used for purposes 

that support the broader aims and welfare of society. As the extent to which the 

number and severity of road accidents recorded in national databases 

represents reality comes under greater scrutiny, the need for this type of study 

will increase. 

While the concept of linkage as described in Section 2.2 is straightforward, a 

variety of approaches was adopted by the partners. It was originally envisaged 

that there would be an exploratory phase of the project where these could be 

compared and harmonised if appropriate. Resource and time constraints meant 

that this was not possible, but it would certainly be important to include this in 

any future research of this type. 

Nevertheless, the various linkage approaches were applied rigorously and all 

work in similar ways using the same variables to identify potential matches, so 

there is no reason to suppose that the results would have differed significantly if 

a common, optimised technique had been applied in all studies. 

As indicated by Table 1, the extent of the eight studies varies widely in time and 

space: from the whole of the Netherlands from 1997-2003 to the Czech town of 

Kromeriz in 2003–05. Similarly, the size of the combined datasets varies widely, 

from 1.6 thousand records from the Czech study to 201 thousand records from 

Scotland. Summary details of the studies are shown in Table 29. 




Linked police and hospital 

records, 

by police severity 

Table 29: Summary of linking results. 

Hospital 

not 

police 

records 

Police not hospital records 

Neither 

police 

nor 

hospital 

records 

Total 

Fatal Serious Slight Unkn Fatal Serious Slight Unkn 

AU 104 2,882 1,689 1,382 12,010 854 5,325 39,800 5,187 69,233 

CZ 41 217 575 129 687 1,649 

FR 3,647 17,616 56,479 1,215 11,500 90,457 

GR 240 1,022 10,005 102 546 11,915 

HU 3 174 323 794 83 485 1,597 3,459 

NL 48,735 11,497 63,353 3,205 2,826 129,616 

ES 13 326 154 1,143 234 1,283 4,933 27 8,113 

UK 427 17,434 8,764 20,672 2,544 12,831 138,334 201,006 

It is inevitable that the strength of the results achieved by various studies differs, 

certainly on statistical grounds and also potentially on the degree to which 

results from a part of a country may be nationally representative. Overall, 

however, the results achieved represent an important step forward in comparing 

the numbers of road accidents and casualties across a range of countries. 

4.1 MAIS and LoS compared 

One of the main aims of Task 1.5 is to recommend a definition of ‘serious 

casualty’ for use in international comparisons of casualty data from the CARE 

database. When the common methodology was being defined, it was seen that 

the broad choice lay between a definition based on the Length of Stay of road 

accident casualties in hospital, and a definition based on their MAIS scores. 

Medical authorities tend to be critical of the use of Length of Stay as an indicator 

of injury severity, e.g. Brasel et al (2007). To the medical layman, it certainly 

appears that Length of Stay is likely to be influenced far more by clinical 

practices and the availability and organisation of hospital services than by the 

level of road safety. It appears that results based on MAIS are more likely to 

monitor casualty and severity trends reliably than results based on Length of 

Stay. 

Results from the Scottish linkage study provide information that is highly 

relevant to this choice of basis. The trends in the linked data between 1980 and 

2005 show how MAIS and Length of Stay for road accident casualties have 

developed over a quarter of a century. The operational procedures for SHIPS 

were unchanged between 1997 and 2005, and indeed for many years 

previously, so any changes in the annual data cannot result from changes in the 

SHIPS data collection procedure. They must be caused by changes in the 

number and nature of casualties, or in the criteria used to admit, treat and 

discharge hospital in-patients. 




Figure 4 compares the distribution of Length of Stay in the linked casualty data 

with three ranges: 0 days (admitted and left hospital on the same day), 1-3 days 

and over 3 days. Note that the y-axis scales differ. This uses the definition of 

Length of Stay that was applied in the earlier study (Keigan et al, 1999) for 

consistency, so the data from the current study have been recalculated using 

these categories. There are clear overall trends, with a shift towards shorter 

stays in hospital. 

Figure 4: Distribution of Length of Stay in Scottish linked casualty data 

20% 

Length of Stay = 0 days 

15% 

10% 

5% 

Car occupant 

Motorcyclist 

Pedal cyclist 

Pedestrian 

All 

0% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

80% 

Length of Stay = 1-3 days 

60% 

Length of Stay >3 days 

70% 

50% 

60% 

40% 

50% 

30% 

40% 

20% 

30% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

10% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

Figure 5 presents the corresponding comparison for MAIS. This comparison is 

affected by the switch from ICD9 to ICD10: SHIPS used the ICD9 system until 

1996, while the ICD10 system was introduced in 1997 and a new mapping from 

ICD to MAIS had to be adopted. The Figure shows major increases in the 

proportion of casualties with MAIS 1 between 1991-95 and 1997-99, and 

corresponding reductions with higher MAIS. It is clear that the combination of 

the ICD10 codes and the new mapping has tended to yield lower MAIS scores. 




60% 

50% 

40% 

30% 

20% 

10% 

Figure 5: Distribution of MAIS in Scottish linked casualty data 

MAIS = 1 

MAIS = 2 

80% 

70% 

60% 

50% 

40% 

30% 

0% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

20% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

35% 

MAIS = 3 

30% 

25% 

20% 

15% 

10% 

Car occupant 

Motorcyclist 

Pedal cyclist 

Pedestrian 

All 

5% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

10% 

MAIS = 4-6 

8% 

6% 

4% 

2% 

0% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

The ICD10 system is likely to provide less accurate estimates of injury severity 

than ICD9, for the reasons discussed in Section 2.3. Nevertheless, provided 

that the bias in the MAIS values calculated from the ICD10 codes is consistent 

between countries and years, it should still provide a valuable benchmark for 

international comparisons of road accident data. 

The Figure does indicate that results based on mapping ICD9 codes to MAIS 

are not comparable with results based on mapping ICD10 codes. Before and 

after the switchover, however, the trends show a consistent pattern which is 

likely to reflect changes in road safety rather than in external influences. 

Unfortunately, Table 1 shows that both systems have been used in the national 




studies. Section 4.4 adjusts the ICD10-based conversion factors to ICD9 in an 

attempt to achieve consistency. 

4.2 Definition of ‘hospitalised’ person 

The title of the final stage of Task 1.5 in the SafetyNet programme of work is 

“Adoption of the common definition of hospitalised persons”. The underlying aim 

of the Task is to achieve a definition of a serious casualty that could be applied 

in all countries. At the time the programme was being developed, it seemed that 

the most likely outcome would be a definition such as “spent more than 3 days 

in hospital”: hence the use of the term hospitalised. 

Overall, however, the evidence from the Scottish linkage study supports other 

arguments against basing the new definition on Length of Stay, and in favour of 

adopting a definition based on MAIS. It would be perverse, then, to adopt a 

definition based on Length of Stay simply because of the terminology used in 

2003 when preparing the programme. 

The only remaining decision concerns the MAIS range to choose for the 

definition of serious casualty. In principle, the threshold could be taken as 2, 

since AIS 2 describes a moderate injury and indeed there are appreciable 

numbers of cases of casualties who die with MAIS=2. The development of the 

estimation procedure in Section 3, however, explains that it is not possible to 

estimate MAIS 1 and 2 separately with the data available in some countries, so 

the minimum feasible value for the threshold is 3. Coincidentally, the AIS 

documentation refers to 3 as a serious injury. The conversion factors from the 

national studies show that adopting a higher threshold would yield rather small 

numbers of serious casualties. 

Recommendation 

Accordingly, it is concluded that the optimal definition of serious casualty for use 

with the CARE database should be a non-fatal casualty with MAIS between 3 

and 6 (inclusive). 

4.3 Trends in the Conversion Factor 

The results of Figure 5 from the Scottish linkage between 1980 and 2005 show 

clear trends in the MAIS distribution over time, and the detailed results from the 

matching process in Section 7.8 show that the relationship between the casualty 

data recorded by the police and by the hospitals is also dynamic. Thus, it is 

likely that Conversion Factors calculated annually would differ from the 

averages for 1997-2005 that were presented in Section 3.1. This is confirmed 

by Figure 6, which presents the annual factors used to estimate the number of 

casualties with MAIS>=3 from the number of serious and slight casualties in the 

STATS19 data. The factors for converting slight casualties are less than for 

serious, so use the right-hand scale. 




Figure 6: Annual MAIS-based Conversion Factors in Scottish linked 

casualty data 

0.30 

0.009 

0.25 

0.20 

0.006 

serious 

0.15 

slight 

0.10 

0.05 

>=3 (serious) 

1997-2005 average(serious) 

>=3 (slight) 

1997-2005 average(slight) 

0.003 

0.00 

0.000 

1997 1998 1999 2000 2001 2002 2003 2004 2005 

Although the annual results are more susceptible to random variation than the 

results for nine years of data, it is clear that there are trends in the factors. 

Ideally, the conversion factors should be recalculated regularly to take account 

of potential variations in the casualty data and their relationship with the medical 

data. 

4.4 Adjustment of ICD10-based results 

It was seen in Section 4.1 that MAIS scores based on ICD10 data are not 

comparable with MAIS scores based on ICD9 data, and that ICD9-based scores 

are likely to be more reliable. Table 1 shows that the Czech and UK national 

studies have used ICD10 data while three used ICD9 (the direct coding system 

of the French and Hungarian studies is comparable with ICD9). Hence, it is 

preferable to adopt ICD9 for the comparison, and a method is needed to adjust 

the Czech and UK results to be comparable with ICD9-based results. 

As yet there has been no study to compare the effect on severity trends of 

changing from ICD9 to ICD10, so the only feasible approach is to make use of 

the trends in the SHIPS/STATS19 data. Figure 7 extends Figure 5 to show the 

proportion of linked casualties with MAIS>=3. A simple statistical model has 

been used to extrapolate the 1980-95 data and estimate the proportions that 

would have been found in 1997-99 if ICD9 had been used. For each road user 

group, the estimated proportion would have been greater than the actual 

proportion and this would have continued through the following years, so the 

conversion factors in Table 8 for MAIS>=3 would have been greater if ICD9 had 

been used. Table 30 shows the adjustments necessary, and applies these to 

the factors from the last row of Table 8. 




Figure 7: Proportion of Scottish linked casualties with MAIS>=3 

MAIS >= 3 

35% 

30% 

25% 

20% 

15% 

10% 

Car occupant 

Motorcyclist 

Pedal cyclist 

Pedestrian 

All 

5% 

0% 

1980-82 1983-85 1986-90 1991-95 1997-99 2000-02 2003-05 

Table 30: Proportion of casualties in 1997-99 with MAIS>=3 

Actual Estimated Adjustment Conversion Factors 

(ICD10-based) (ICD9-based) to ICD9 Serious Slight 

Car occupant 12.1% 24.7% 2.03 0.30 0.008 

Motorcyclist 18.1% 35.4% 1.95 0.52 0.022 

Pedal cyclist 9.3% 22.1% 2.38 0.69 0.021 

Pedestrian 18.5% 21.2% 1.14 0.26 0.009 

Other 0.30 0.007 

All 14.1% 24.5% 1.74 0.34 0.009 

Applying the overall adjustment to the Czech conversion factors (Table 13) 

gives 0.37 for serious casualties and 0.031 for slight. 

The overall results from the national studies are brought together in the 

following two Figures. First, Figure 8 presents the conversion factors for 

MAIS>=3, with the factor for slight casualties multiplied by 10 (the Czech and 

UK factors have been adjusted to ICD9). For example, from Table 16, for every 

serious casualty in the French accident data there are 0.68 casualties with 

MAIS>=3, while for every slight casualty there are 0.06. 




Figure 8: Conversion Factors for MAIS>=3, all road users 

Conversion Factors 

1.0 

0.5 

N1 N2*10 

0.0 

France Hungary Greece Netherlands 

Czech 

Republic 

UK 

Spain 

Next, Figure 9 compares the conversion factors for Length of Stay>=1 day, 

distinguishing between 1-3 and >3 days. Even the ranking of countries in the 

two Figures is different. 

Figure 9: Conversion Factors for Length of Stay>=1 day, all road users 

1.5 

Conversion factors 

1.0 

0.5 

1-3 days >3 days 

0.0 

Serious 

Slight 

Serious 

Slight 

Serious 

Slight 

Serious 

Slight 

Serious 

Slight 

Serious 

Slight 

Serious 

Slight 

Serious 

Slight 

Greece Netherlands Austria France UK Spain Hungary Czech 

Republic 

The clear differences shown in these Figures between countries confirm that 

casualty reporting practices differ markedly, and that it would be misleading to 

compare national casualty data without adjustment. The differences 




demonstrate the need for the use of conversion factors such as those presented 

here. 

4.5 An application 

The conversion factors presented in Section 3 (adjusted to ICD9 basis as 

necessary) will now be applied using actual CARE data, to illustrate the ways in 

which the scope for international comparisons is expanded by the results of this 

research. Naturally, the strength of the results depends upon the assumption 

that the factors are nationally representative. 

Table 31 presents the 2003-05 average annual casualty totals from CARE. At 

the time the data were downloaded, no post-2003 data were available for the 

Netherlands, only data from 2005 were available for Hungary and no data were 

available for the Czech Republic. The definition of serious and slight casualty in 

France changed in 2005, so the conversion factors only apply up to 2004. 

Table 31: 2003-05 annual average casualty totals 

Killed Serious Slight 

Austria 859 13,956 41,367 

France (2002-04) 6,414 19,898 100,587 

Greece 1,644 2,338 18,650 

Hungary (2005) 1,278 8,320 19,185 

Netherlands (2001-03) 1,003 10,881 29,608 

Spain 4,861 23,323 117,286 

UK 3,454 32,445 254,253 

Source: CARE database, June 2007 

To distinguish the number of serious casualties estimated according to the 

definition proposed above, i.e. with MAIS between 3 and 6, these will be 

referred to as serious* casualties. The conversion factors illustrated in Figure 8 

are applied in Table 32 to estimate the number of serious* casualties, showing 

the components N1 and N2. The relation between the number of serious and 

serious* casualties varies widely. In Greece, the estimated number of serious* 

casualties considerably exceeds the number of serious casualties, while in the 

Netherlands, Spain and the United Kingdom it is less than one half. 




Table 32: Estimation of the number of Serious* casualties 

Serious casualties Slight casualties 

CARE 

total 

factor 1 N1 CARE 

total 

factor 2 N2 Serious* 

=N1+N2 

Serious* 

Serious 

France 19,898 0.68 13,612 100,587 0.061 6,157 19,768 0.99 

Greece 2,338 0.46 1,081 18,650 0.121 2,259 3,339 1.43 

Hungary 8,320 0.48 3,962 19,185 0.040 761 4,723 0.57 

Netherlands 10,881 0.39 4,254 29,608 0.016 474 4,728 0.43 

Spain 23,323 0.26 6,084 117,286 0.018 2,059 8,143 0.35 

UK 32,445 0.34 11,130 254,253 0.009 2,298 13,428 0.41 

In order to allow for the size of country, fatality rates per million population are 

often compared, also fatality rates per million motor vehicles. Table 33 

compares these fatality rates with the rates of serious* casualties. IRTAD 

counts of population and motor vehicles in 2005 have provided the 

denominators for these calculations. 

Table 33: Casualty rates per million population and per million motor 

vehicles, 2003-05 

Rate per million 

population 

Rate per million 

vehicles 

Killed Serious* Killed Serious* 

France (2002-04) 106 326 173 532 

Greece 148 301 248 503 

Hungary (2005) 127 468 379 1,402 

Netherlands (2001-03) 61 290 116 548 

Spain 112 187 176 294 

UK 57 223 103 399 

These results need to be qualified in several ways. The main qualification is that 

some conversion factors may not be nationally representative, principally 

because of the choice of study area. In addition, the data for the Netherlands 

come from an earlier period than for the other countries as the post-2003 Dutch 

fatality data were not available in CARE in June 2007. It is known that the 

national casualty totals have fallen since 2003 so the actual rates for the 

Netherlands for 2003-05 are lower than shown in the table. 

Also, no account has been taken of trends in the conversion factors. In view of 

the results in the previous section, it might seem better to choose a period for 

each country that was centred on the period of the national study. This would 

lead to greater inconsistency among the periods being compared, and introduce 

a different form of bias. The approach adopted above seems preferable, but it 

will be important to take account explicitly of trends in the conversion factors in 

any future study of this general type. 




5. Conclusions and Recommendations 

At present, international comparisons of the level of road safety rely almost 

exclusively on the analysis of data relating to fatal accidents and fatalities. This 

is unsatisfactory, for by most criteria non-fatal accidents and casualties impose 

a burden on society that is at least as great. It is potentially misleading to 

analyse the level of road safety in a country only in terms of fatal accidents and 

fatalities. 

The CARE database contains details of non-fatal accidents and casualties, so it 

is essential to allow these data to contribute to international comparisons of 

road safety. The problems with doing this have been well known for many 

years, and scientifically acceptable methods are needed to overcome these 

problems. The studies presented in this report represent an important step in 

this direction. They do not offer a complete solution since they only deal with the 

more seriously injured casualties, namely those that attend hospital for 

treatment. Nevertheless, this is the most important section of the spectrum of 

non-fatal casualties, and the underlying principles may be developed in future to 

extend the coverage. The results that have been achieved demonstrate that the 

problems mentioned above are not simply theoretical but are real and acute. 

The original objective for this project included those casualties who were slightly 

injured, but the resources available have meant that no practical progress has 

been made in this direction. At present, police reports are the only systematic 

source of information about those with less serious injuries, although subject to 

the limitations that have been examined in this report. An independent source 

would be required in order to assess the completeness of these reports. Such 

data collection is expensive, and has not been attempted within this project. 

Nevertheless, approaches exist that could be considered for future research. 

One possibility would be to include questions relating to road accident in the 

EC‘s SARTRE survey. The relatively low incidence of road accidents in much of 

Europe may rule this out on statistical grounds, but it would be possible to 

include questions in the regular national surveys carried out by many 

Governments, such as the General Household Survey in Great Britain. There is 

relevant experience in the Netherlands (AVV, 2002). 

Various technical problems have been encountered and surmounted in the 

course of this project. It is probably too early to say that the coefficients 

estimated from these linkage studies allow the national casualty data from the 

participating countries to be compared in an unbiased fashion. Nevertheless, 

they provide important new information and demonstrate clearly how fully 

reliable and nationally representative coefficients may be prepared. 

While the national studies have been carried out for a specific purpose within 

the SafetyNet project, the linked data sets have a research potential that 

extends far beyond this purpose and that has not been considered in this report. 




Recommendations 

1. The definition of a “serious” casualty for use with the CARE database 

should be a non-fatal casualty with MAIS between 3 and 6 (inclusive). At 

present this can only be implemented in the countries reported here (other 

than Austria). It is recommended that a new study be carried out so that this 

definition can be implemented in more countries. Naturally, this study would 

take into account the lessons learned in the study reported here. 

2. This approach applied in this project represents an important step forward in 

comparing the numbers of road accidents and casualties across a range of 

countries. In view of the variation of the Conversion Factors among the 

countries studied, however, it would not be wise to generalise results to 

other countries. A new study is needed that includes more Member States. 

Also, where the current study included only part of a country, the coverage 

within that country should be extended where possible. 

3. The CODES system of US Federal Government’s National Highway Traffic 

Safety Administration offers an example of how this might be achieved. The 

NHTSA routinely supports the state-wide linkage of accident and medical 

records in about 30 States. The EC could support regular national linkage 

studies in the Member States. The results would have many benefits in 

addition to the preparation of conversion factors for use with CARE. 

4. Police accident reporting practices in any country will evolve over time, so 

the relationship between the police casualty statistics and the actual number 

of casualties is also likely to change. Clear evidence of this has been seen 

in both studies that examined developments over time (the Netherlands and 

Scotland). It will be necessary to repeat the linkage studies regularly in 

order to update the conversion factors. 

5. The methods used to link police and medical records in the national studies 

differ in detail, as described in the reports in the Appendix. Some 

differences were inevitable since the levels of detail in the datasets being 

linked differed from country to country. Other differences, however, reflected 

differences in background and experience among those carrying out the 

linkage. Ideally, each of the various methods would have been applied to 

each dataset to see whether the linkages achieved depended significantly 

upon the method used, and if so to identify the optimal method. 

Unfortunately the resources available for this study did not allow for this 

comparative phase to be carried out, but it would undoubtedly be valuable 

for any future study of this type to incorporate such a comparison of the 

methods available among the partners. 

6. The methods used to estimate MAIS from ICD injury codes need better 

validation. It is not necessary or desirable to confine this to samples of road 

accident casualties, a broad application to all injuries however caused 

would also be very helpful. 




7. It would be valuable to estimate the number of casualties missed by both 

the police and the hospital reporting, in order that the conversion factors can 

fully account for under-reporting. The number of these casualties can be 

estimated by the capture-recapture method, which requires certain 

assumptions to be met. In particular, the fact that the police under-reporting 

rate is associated with characteristics such as injury severity and mode of 

transport must be taken into account. This can be done by stratification, in 

other words estimating the number of missed casualties (and hence the 

total number of casualties) in each strata defined by these characteristics. 

The conversion factors should not only be estimated according to injury 

severity and mode of transport (as done here) but also according to other 

relevant characteristics such as number of vehicles involved and type of 

road (e.g. urban/rural). Moreover, if we estimate conversion factors using 

regional data and wish to apply them nationally, the need to adjust for 

urban/rural characteristics is greater. 

8. This study has linked police and hospital records, so inevitably can say 

nothing about those road accident casualties with lesser injuries that do not 

attend hospital for treatment. While individually these people are less 

severely injured than those who attend hospital, they are likely to be more 

numerous. It will be important to develop techniques that can prepare 

conversion factors for this group of casualties. The new IDB (Injury 

Database) of DG-Sanco will contain detailed information about traffic 

accidents in the near future from samples of accidents across the EU. This 

new source of information could potentially be used in linkage studies. 




6. References 

Association for the Advancement of Automotive Medicine (1990). The 

Abbreviated Injury Scale (1990 revision), Des Plaines, Illinois. 

AVV, Ministerie van Verkeer en Waterstaat, (2002). Verkeersongevallen in 

Nederland 2001, Heerlen, The Netherlands. 

http://www.rws-avv.nl/pls/portal30/docs/3736.PDF 

Brasel K J, Lim H J, Nirula R and Weigelt J A (2007). Length of Stay: an 

appropriate quality measure Archives of surgery, vol 142, pp 461-466. 

Crash Outcome Data Evaluation System (CODES). 

http://www-nrd.nhtsa.dot.gov/departments/nrd-30/ncsa/codes.html,, visited 25 

June 2007 

Department of Health and Human Services. Health Resources and Services 

Administration. From E to VWXY Cause of Injury Codes. 

ftp://ftp.hrsa.gov/mchb/vwxy.pdf 

European Centre for Injury Prevention, University of Navarra (2006). 

Algorithm to transform ICD-10 codes AIS and ISS, version 1 for SPSS. 

Pamplona, Spain 

Expert Group on Injury Severity Measurement (EGISM) (2004). Discussion 

document on injury severity measurement in administrative datasets. 

http://www.cdc.gov/nchs/data/injury/DicussionDocu.pdf. 

Lopez D G, Rosman D L, Jelinek G A, Wilkes G J and Sprivulis P C (1999). 

Complementing police road-crash records with trauma registry data – an initial 

evaluation. Accident Analysis and Prevention, Vol. 32, pp. 771-777. 

MacKenzie, E. J., Sacco, W et al. (1997). ICDMAP-90: A users guide. 

Baltimore, The Johns Hopkins University School of Public Health and Tri- 

Analytics, Inc. 

National Center for Health Statistics (2007). International Classification of 

Diseases 10th Revision (ICD-10). 

http://www.cdc.gov/nchs/about/major/dvs/icd10des.htm. 

Osler, T., Rutledge, R., et al. (1996). ICISS: an international classification of 

disease-9 based injury severity score. J. Trauma 41 (3), 380–388. 

Stephenson S, Langley J and Cryer C (2005). Effects of service delivery 

versus changes in incidence on trends in injury: a demonstration using 

hospitalised traumatic brain injury. Accident Analysis and Prevention 2005; 

37(5):825-832. 

World Health Organization (1992). International Statistical Classification of 

Diseases and Related Health Problems, Tenth Revision. Geneva 




7. Appendix A 

This Appendix includes eight reports that present details of the national studies 

whose results appeared in the main report. Each report was prepared by the 

person responsible for carrying out that national study. The reports have a 

shared structure, but are formatted independently. 

7.1 Study carried out in Austria 

Report prepared by Stefan Hoeglinger (KfV) 

7.1.1 Introduction 

The Austrian national study combined the police database of accidents 

throughout Austria in the year 2001 with the Austrian hospital discharge 

database. 

7.1.2 Description of data sources 

The police database 

The police database of Austria comprises data of road accidents on public 

roads with at least one person injured. This statistics should cover 100% of all 

relevant cases in Austria. The police collect the data by filling in a standardised 

form. All these papers are collected by the Federal Statistics body (Statistik 

Austria) and transformed into an electronic format. Every month, the Federal 

Statistics body is sending out the monthly data of road accidents. Persons 

committing suicide in a road traffic accident are not included in this database. If 

the probability is very high that the person died due to a serious health problem 

(e.g. heart attack) the respective record will also be removed from the database. 

For the linking procedure and the preparation of matrixes 1 and 2 accident data 

of the year 2001 are used. In this year 57.223 persons were either injured or 

killed in a road accident. In contradiction to other European countries the 

Austrian police use the injury type “injury unknown” when the police agent on 

the scene does not know if the person is slightly or seriously injured. Within the 

CARE system these persons are generally treated as seriously injured. Like 

most European countries, Austria is using the 30 day definition for fatalities: If a 

person dies within 30 days after the accident he or she will be counted as a 

road accident fatality. Uninjured persons involved in the accidents are excluded 

from this linking process. 

The hospital discharge database 

In the hospital discharge database administrative and medical data of all inpatients 

of 270 Austrian hospitals are collected. The hospital discharge records 

are designed to fulfil the needs of financial compensation for medical services. 




This statistics should cover 100% of all relevant cases in Austria. Out-patients 

are not recorded in this database. Data is collected by the hospitals and 

transferred on an electronically basis to the Ministry for Health and Women. The 

ministry passes the data to the Federal Statistics body (Statistik Austria) which 

prepares the so called “Spitalsentlassungsstatistik” (hospital discharge 

statistics). 

For our needs we analysed all in-patients injured in road accidents, 

corresponding to ICD 10 coding (S00-T99) and the external cause of injury “not 

at work” or “at work”. In this database only the main diagnosis at the moment of 

discharging from the hospital is recorded. Before 2001, the hospitals reported 

the main diagnosis using the ICD 9 coding system. 

The hospital discharge database is case-orientated instead of personorientated. 

If a person is transferred from one hospital to another he or she is 

recorded twice. It is thus not possible to recognise that he or she is just a single 

person. As the hospitals collect data about the patient’s age when leaving the 

hospital, the age at the moment when admitted to the hospital has to be 

recalculated by using the date of birth and the date of admission to the hospital. 

The hospital discharge data is published on an annual basis. Due to some late 

transmission of data by hospitals, there is a delay of 1½ years between the year 

of discharge and the year of data publication. Every year, selected indicators 

from this database are calculated and sent to EUROSTAT, WHO-HFA (Health 

for All) and also to the OECD database (OECD – Health Data). 

In order to prepare data for the linking procedure, data from 2001 and 2002 

were used in order to prepare a database with recorded persons who were 

admitted to a hospital in 2001. Only records in the hospital discharge database 

with an indication of a road traffic accident (“not at work” or “at work”) where 

used for the linking procedure. (IX14_Verletzungsg; Cause of Injury ISIS S30) 

Without this limitation around 250,000 records of injuries and toxications are in 

the database each year. When applying this limitation only 18,067 records are 

remaining. 

7.1.3 Description of the linking process 

Before the linking procedure is started, the data files have to be properly 

prepared. All this was done by using the standard software SPSS and MS 

Access. 

Variables used for the linking process 

Table 34 shows the key variables used in the linking process. 




Common 

name 

Date 

Table 34: Key variables in the police and hospital databases 

Hospital discharge database 

Police database on road accidents 

Explanation Variable Explanation Variable 

Date when admission to 

Date when the 

the hospital (no 

accident happens 

H_Date 

unknown in the 

(no unknown in the 

P_Date 

database) 

database) 

Hour of the day when 

the accident happens 

-0h-21h and 22h-24h P_Hour 


- - - 

Sex 

Austrian or 

Foreigner 

Age 

Federal 

state 

Living 

federal 

state 

Gender (no unknown in 

the database) 

Nationality of person 

(derived from the 

passport not from the 

country of living) 


database) 

Age (no unknown in the 

database) 

Federal state where the 

hospital is located. (no 


database) 

Federal state where the 

person lives. (no 


database) 

H_Sex 

H_Foreigner 

H_Age 

H_Fed 

database) 

Gender (no unknown 

in the database) 

Nationality of person 

(derived from the 

passport not from the 

driving license or 

license plate; no 


database) 

Age (24 records with 


database) 

Federal state where 

the accident 

happened. (no 


database) 

H_Fed - - 

P_Sex 

P_Foreigner 

P_Age 

P_Fed 

As mentioned above, in the police database only records of injured or killed 

persons are documented and from the hospital discharge database only records 

of persons involved in a road accident are used. This pre-selection is necessary 

because the variables for the linking process are very limited. Otherwise the 

chance of having a random (and therefore more likely wrong) match between 

records of both databases would be relatively high. 

Technical preparation 

All the variables used (Table 34) are prepared to use common values and 

variables. For the linking process the following additional variables are added to 

the databases. 




Table 35: Additional variables used for the linking process between the 

police and hospital databases 

Police database 

P_Nr 

P_Lsi 

P_Pointer1 

P_Dist1 

P_Pointer2 

P_Dist2 

P_Matchnr 

P_Selectivity 

P_Distfinal 

Hospital database 

H_Nr 

H_Lsi 

H_Pointer1 

H_Dist1 

H_Pointer2 

H_Dist2 

H_Matchnr 

H_Selectivity 

H_Distfinal 

First, the databases were arranged in ascending order of the variable H_Date 

and P_Date. After that, the records were consecutively numbered using the 

variable H_Nr and P_Nr. Later on these numbers are refered to as the rank 

numbers of the record. 

P_Lsi and H_Lsi indicate the status of the record. The values for this variable 

are: 

-2 died on the scene 

-1 undecided 

0 matched 

1 not matchable 

2 matchable but not matched due to more than one perfect fits. 

“Double zeros” 

The variable P_Dist1 and H_Dist1 is used to calculate the distance to the best 

neighbour of a record. The variable P_Dist2 and H_Dist2 contain the distance to 

the next best neighbour of the relevant record. The variable P_Pointer1 and 

H_Pointer1 refer to the rank number of the best neighbour. The variable 

P_Pointer2 and H_Pointer2 refer to the rank number of the next best neighbour. 

P_Selectivity and H_Selectivity show the calculated final selectivity between 

matched pairs. P_Distfinal and H_Distfinal show the final distance between 

matched records. 

Preparation of the record set of the hospital admission database 

For the analysis it is necessary to derive a database which contains records of 

those persons who were admitted to the hospital in the year 2001. As the 2001 

hospital discharge database contains only records of persons who left the 

hospital in the year 2001, all records from the hospital discharge databases of 

the years 2001 and 2002 with an admission date in the year 2001 were 

selected. The result was that 233 persons were admitted to the hospital in 2001 

and left the hospital in 2002. It can be assumed that there are just a few cases 

of people who were admitted to the hospital in 2001 and left the hospital after 

the year 2002. This was thus not analysed. 




We presume that a very high percentage of all relevant road accident victims 

get admitted to a hospital within 4 days after the accident. Therefore the 136 

hospital admission database records until 4 January 2002 where included as 

well. 

The variable “Age” of the hospital discharge database contains the age of the 

person when leaving the hospital. Therefore the variable, H_Age (age at the 

moment of hospital admission) is calculated by using this variable “Age”, the 

date of admission to the hospital and the date of birth. 

The Linking procedure 

The attempt to match the Austrian police’s road traffic accident database with 

the hospital discharge (admission) database used a procedure very similar to 

the one proposed by SWOV in 2001. 

The police database for 2001 contains 57.223 road traffic accidents, whereas 

the hospital database contains 18.067 hospitalisations. By simply joining every 

record of the police database with every record of the hospital database about 1 

billion matches would have to be assessed. By using a slightly modified 

procedure as the one proposed by SWOV the assessment is limited to a 

reasonable amount. The quality of the linking process has been improved 

steadily by running iterative series of tests. 

The similarity of the matches between records of both databases is calculated 

by the distance function which is described in “The distance function”. The 

quality or uniqueness of the match is calculated by the selectivity function. This 

is explained in “The selectivity function”. In “The linking procedure” the linking 

procedure itself is described. 

The distance function 

The personal ID-number, which could serve as a primary key, is not recorded in 

the Austrian police and hospital reports, so a set of other characteristics has to 

be used to match the respective records. These key variables (see Table 34) 

are used in the distance function. 

When comparing the values of the key variables they are sometimes not 

correctly registered or even missing. Furthermore, the time span between the 

occurrence of the accident and the time of admission to the hospital depends on 

some parameters. It is uncertain that a seriously injured person is brought to a 

hospital with a delay of two days. But it is more probable that a slightly injured 

person might have some problems two days after the accident and then go to 

the hospital. To quantify the similarity between two records of the police and 

hospital databases a generalised distance function has been defined by SWOV. 

A very low distance close to zero indicates a very high probability that the 

person in the police database is the same as the one recorded in the hospital 




database. If the distance is higher (because the key variables are different in 

some variables) the probability that this matched pair refers to the same victim 

is smaller. 

As mentioned by SWOV (2001) the distance function can be described as 

follows: 

“Let the hospital database contain N 1 records and the police database 

contain N 2 records, and let c ik denote the category of record i on key 

variable k (k = 1,…., m), then the distance between record i (i=1, …, N 1 ) 

and record j (j = 1,…, N 2 ) is defined as: 

m 

d ij = ∑ δ ( c c ik 

, ). 

(1) 

k = 1 

jk 

Very generally, the term δ c ik 

, c ) 

( 

jk 

⎧ 0 if c 

⎫ 

ik 

= c 

jk 

⎪ 

⎪ 

δ ( cik 

, c 

jk 

) = ⎨ ak 

if cik 

≠ c 

jk ⎬ 

(2) 

⎪ 

⎪ 

⎩bk 

if cik 

and / or c 

jk 

missing⎭ 

Although the values of a k and b k in (2) are defined for each key variable, 

they all have in common that they increase the distance between two 

records when the records contain unequal categories and/or missing 

information on a key variable. “ 

The determination of a k and b k for the following key variables was ruled by the 

assumption that a distance of 100 corresponds to a probability of about 50% 

that two records refer to the same victim (see Figure 10). The determination of 

the distances is a weakness of the used methodology. By carrying out several 

trials the distances were chosen manually. 




Figure 10: Determination of distances: Probability – Distance function 

Probability - Distance function 

Probability 

100% 

90% 

80% 

70% 

60% 

50% 

40% 

30% 

20% 

10% 

0% 

0 100 200 300 400 500 

Distance 

Table 36: Values for a k for the key variable “Date” 

P_Inj_Name P_Hour_Name Date_Diff ak P_Inj_Name P_Hour_Name Date_Diff ak 

Slightly injured before 21 0 100 Killed within 24 hours after 21 2 1000 

Slightly injured after 21 0 100 Killed within 24 hours before 21 3 1000 


Slightly injured after 21 1 110 Killed within 24 hours before 21 4 1000 


Slightly injured after 21 2 150 Killed between 24 and 48 hours before 21 0 0 

Slightly injured before 21 3 200 Killed between 24 and 48 hours after 21 0 0 


Slightly injured before 21 4 300 Killed between 24 and 48 hours after 21 1 20 


Injury unknown before 21 0 0 Killed between 24 and 48 hours after 21 2 1000 

Injury unknown after 21 0 0 Killed between 24 and 48 hours before 21 3 1000 









Seriously injured before 21 0 0 Killed between 48 and 72 hours after 21 2 1000 

Seriously injured after 21 0 0 Killed between 48 and 72 hours before 21 3 1000 


Seriously injured after 21 1 20 Killed between 48 and 72 hours before 21 4 1000 


Seriously injured after 21 2 1000 Killed between 72 and 30 days before 21 0 0 

Seriously injured before 21 3 1000 Killed between 72 and 30 days after 21 0 0 


Seriously injured before 21 4 1000 Killed between 72 and 30 days after 21 1 20 


Killed within 24 hours before 21 0 0 Killed between 72 and 30 days after 21 2 1000 

Killed within 24 hours after 21 0 0 Killed between 72 and 30 days before 21 3 1000 


Killed within 24 hours after 21 1 20 Killed between 72 and 30 days before 21 4 1000 


In Table 36, P_Inj_Name is the severity of injury reported in the police 

database. P_Hour_Name indicates if the accident happens between 0h and 21h 

or between 22h and 24h. Date_diff is the number of days between the date 




when the accident happened and the date where the patient was admitted to 

the hospital. 

b k : n/a as no date is missing in one of the databases 

Table 37: Values for a k for the key variable “Sex” 

P_Sex_Name H_Sex_Name ak 

male male 0 

male female 2000 

female male 2000 

female female 0 

Where P_Sex_Name is the gender information in the police database and 

H_Sex_Name is the gender information in the hospital file. 

b k : n/a as no gender information is missing in one of the databases 

Key variable: Austrian or Foreigner 

a k : a k = 0 if the value of this variable is the same in both databases 

b k : n/a as no information is missing in one of the databases 

If the value of this variable differs in the databases, the match is treated as not 

linkable at all. 

Key variables about the location in Austria 

The following pages describe key variables related to federal states in Austria. 

Figure 11 shows a map of Austria including all federal states to get an 

impression about the chosen a k . 

Figure 11: Map of Austria 

Niederoesterreich 

Wien 

Oberösterreich 

Vorarlberg 

Tirol 

Salzburg 

Steiermark 

Burgenland 

Kaernten 




Key variable: Federal state 

a k : 

Table 38: Values for a k for the key variable “Federal state” 

P_Fed_Name H_Fed_Name ak P_Fed_Name H_Fed_Name ak 

Burgenland Burgenland 0 Salzburg Steiermark 250 

Burgenland Kaernten 250 Salzburg Tirol 35 

Burgenland Niederoesterreich 125 Salzburg Vorarlberg 500 

Burgenland Oberoesterreich 250 Salzburg Wien 35 

Burgenland Salzburg 500 Steiermark Burgenland 50 

Burgenland Steiermark 125 Steiermark Kaernten 125 

Burgenland Tirol 35 Steiermark Niederoesterreich 125 

Burgenland Vorarlberg 500 Steiermark Oberoesterreich 125 

Burgenland Wien 35 Steiermark Salzburg 125 

Kaernten Burgenland 500 Steiermark Steiermark 0 

Kaernten Kaernten 0 Steiermark Tirol 35 

Kaernten Niederoesterreich 250 Steiermark Vorarlberg 500 

Kaernten Oberoesterreich 350 Steiermark Wien 35 

Kaernten Salzburg 125 Tirol Burgenland 500 

Kaernten Steiermark 125 Tirol Kaernten 125 

Kaernten Tirol 35 Tirol Niederoesterreich 500 

Kaernten Vorarlberg 500 Tirol Oberoesterreich 500 

Kaernten Wien 35 Tirol Salzburg 75 

Niederoesterreich Burgenland 125 Tirol Steiermark 500 

Niederoesterreich Kaernten 500 Tirol Tirol 0 

Niederoesterreich Niederoesterreich 0 Tirol Vorarlberg 250 

Niederoesterreich Oberoesterreich 250 Tirol Wien 35 

Niederoesterreich Salzburg 500 Vorarlberg Burgenland 500 

Niederoesterreich Steiermark 125 Vorarlberg Kaernten 500 

Niederoesterreich Tirol 35 Vorarlberg Niederoesterreich 500 

Niederoesterreich Vorarlberg 500 Vorarlberg Oberoesterreich 500 

Niederoesterreich Wien 35 Vorarlberg Salzburg 500 

Oberoesterreich Burgenland 500 Vorarlberg Steiermark 500 

Oberoesterreich Kaernten 500 Vorarlberg Tirol 35 

Oberoesterreich Niederoesterreich 75 Vorarlberg Vorarlberg 0 

Oberoesterreich Oberoesterreich 0 Vorarlberg Wien 35 

Oberoesterreich Salzburg 250 Wien Burgenland 500 

Oberoesterreich Steiermark 125 Wien Kaernten 500 

Oberoesterreich Tirol 35 Wien Niederoesterreich 500 

Oberoesterreich Vorarlberg 500 Wien Oberoesterreich 500 

Oberoesterreich Wien 35 Wien Salzburg 500 

Salzburg Burgenland 500 Wien Steiermark 500 

Salzburg Kaernten 125 Wien Tirol 35 

Salzburg Niederoesterreich 350 Wien Vorarlberg 500 

Salzburg Oberoesterreich 75 Wien Wien 0 

Salzburg Salzburg 0 

P_Fed_Name indicates the federal state where the accident happened 

(mentioned in the police database) and H_Fed_Name indicates in which federal 

state the hospital is located. 





Key variable: Federal state of living 

a k : 

Table 39: Values for a k for the key variable “Living” 

P_Fed_Name H_Liv_Name ak P_Fed_Name H_Liv_Name ak 

Burgenland Burgenland 0 Tirol Salzburg 70 

Kaernten Burgenland 150 Vorarlberg Salzburg 150 

Niederoesterreich Burgenland 70 Wien Salzburg 150 

Oberoesterreich Burgenland 150 Burgenland Steiermark 70 

Salzburg Burgenland 150 Kaernten Steiermark 70 

Steiermark Burgenland 70 Niederoesterreich Steiermark 70 

Tirol Burgenland 150 Oberoesterreich Steiermark 70 

Vorarlberg Burgenland 150 Salzburg Steiermark 150 

Wien Burgenland 40 Steiermark Steiermark 0 

Burgenland Kaernten 150 Tirol Steiermark 150 

Kaernten Kaernten 0 Vorarlberg Steiermark 150 

Niederoesterreich Kaernten 150 Wien Steiermark 150 

Oberoesterreich Kaernten 150 Burgenland Tirol 150 

Salzburg Kaernten 70 Kaernten Tirol 70 

Steiermark Kaernten 70 Niederoesterreich Tirol 150 

Tirol Kaernten 70 Oberoesterreich Tirol 150 

Vorarlberg Kaernten 150 Salzburg Tirol 70 

Wien Kaernten 150 Steiermark Tirol 150 

Burgenland Niederoesterreich 70 Tirol Tirol 0 

Kaernten Niederoesterreich 70 Vorarlberg Tirol 70 

Niederoesterreich Niederoesterreich 0 Wien Tirol 150 

Oberoesterreich Niederoesterreich 70 Burgenland Vorarlberg 150 

Salzburg Niederoesterreich 150 Kaernten Vorarlberg 150 

Steiermark Niederoesterreich 70 Niederoesterreich Vorarlberg 150 

Tirol Niederoesterreich 150 Oberoesterreich Vorarlberg 150 

Vorarlberg Niederoesterreich 150 Salzburg Vorarlberg 150 

Wien Niederoesterreich 40 Steiermark Vorarlberg 150 

Burgenland Oberoesterreich 150 Tirol Vorarlberg 70 

Kaernten Oberoesterreich 150 Vorarlberg Vorarlberg 0 

Niederoesterreich Oberoesterreich 70 Wien Vorarlberg 150 

Oberoesterreich Oberoesterreich 0 Burgenland Wien 40 

Salzburg Oberoesterreich 70 Kaernten Wien 70 

Steiermark Oberoesterreich 70 Niederoesterreich Wien 40 

Tirol Oberoesterreich 150 Oberoesterreich Wien 70 

Vorarlberg Oberoesterreich 150 Salzburg Wien 70 

Wien Oberoesterreich 150 Steiermark Wien 70 

Burgenland Salzburg 150 Tirol Wien 150 

Kaernten Salzburg 70 Vorarlberg Wien 150 

Niederoesterreich Salzburg 150 All federal states Not Austria 0 

Oberoesterreich Salzburg 70 All federal states Unknown 0 

Salzburg Salzburg 0 All federal states Austria but unknown 0 

Steiermark Salzburg 70 

P_Fed_Name indicates the federal state where the accident happened 

(mentioned in the police database) and H_Liv_Name indicates the federal state 

where the hospitalised person lives. 





Key variable: Age 

a k : a k = 0 if the value of this variable is the same in both databases 


If the value of this variable differs in the databases, the match is treated as not 

linkable at all. 

The selectivity function 

The similarity of the records can be quantified by calculating the distance 

between two records of the two databases. If a record of the police database 

finds a very similar record (small distance) in the hospital database it is 

important to control if there are also other hospital records existing with a small 

distance to the initial record in the police database. In such a case, the 

uniqueness of the initial pair can be criticised. 

The selectivity of a matched pair is the minimum of the differences of the 

distances to its best neighbour and its next best neighbour of each record. 

Whereas the best neighbour is the record in the other database with the lowest 

distance and the next best neighbour is the record in the other database with 

the second lowest distance to the selected current record. 

If the selectivity is high, the uniqueness is high too. If the selectivity is low, it is 

debatable whether the best neighbour pair or the next best neighbour refers to 

the same person. 

The linking procedure 

The linking procedure is implemented by using the standard software MS 

Access 2003. In the following section the method of how to find matches 

between the hospital and the police database is described. The matches should 

have a high probability of referring to the same person in each database. The 

section “first pass” is more or less similar to the paper of SWOV (2001). In the 

section “second pass” an additional function is implemented to cover all 

possible cases of such a linking procedure. 

First initiation 

The first initiation starts after both databases were properly prepared. 

In this initiation the following variables are filled with constant values: 

P_Lsi = -2 or -1 H_Lsi = -1 

P_Dist1 = 100000 H_Dist1 = 100000 

P_Dist2 = 100000 H_Dist2 = 100000 

P_Pointer1 = -1 H_Pointer1 = -1 




P_Pointer2 = -1 H_Pointer2 = -1 

If the severity of injury reported in the police database indicates that the person 

died on the scene the variable concerning the status of the record P_Lsi is set 

to -2 (died on the scene). In all other cases the value of P_Lsi and H_Lsi is set 

to -1 (undecided) 

At the beginning of the linking procedure very high distances are applied to the 

relevant variables. During the linking procedure these variables are recalculated 

and in case of a lower distance the variables are replaced with new values. 

An initial value of “-1” is selected for all pointers. 

First pass 

In the first pass the best neighbour and the next best neighbour is determined 

by using the distance function. 

If a person died on the scene (P_Lsi = -2) the related record will be excluded 

from further examination as it can be assumed that this person will not be 

transferred to a hospital. 

As mentioned above, only records where the admittance to the hospital occurs 

within 4 days after the accident were considered. The first record of the police 

database is compared with all records in the hospital database satisfying this 

time span restriction. When two records are compared it is checked, if any of 

the already stored distances in the variable .._Dist1 is higher as the new 

calculated one. If so, the old distance is updated by the new distance and stored 

in both records. If the new calculated is higher than the distance stored in the 

variable .._Dist1, but lower than the distance stored in the variable .._Dist2, the 

variable .._Dist2 is updated with the new calculated distance and stored in both 

records of the databases. 

Also the corresponding pointers H_Pointer1, H_Pointer2, P_Pointer1, 

P_Pointer2 are updated with the rank number of the relevant records of the 

other database. This procedure is done for both records which are compared. 

After that, the second record of the police database is compared to all records in 

the hospital database also satisfying the mentioned timespan restriction. The 

same systematic treatment is applied. This procedure continues until the last 

record of the police database is processed. 

At the end of this first pass all records which found a similar record in the other 

database contain a pointer which refers to the rank number of its best neighbour 

in the other database and also a pointer which refers to the rank number of its 

next best neighbour in the other database. Each of these neighbours is 

combined with the distance between these records in both databases. 




Second initiation 

In this initiation the following variables are filled with these values: 

P_Matchnr = 0 H_Matchnr = 0 

P_Selectivity = 0 H_Selectivity = 0 

P_Distfinal = 100000 H_Distfinal = 100000 

P_Lsi = 1 or -1 H_Lsi = 1 or -1 

P_Matchnr and H_Matchnr indicate the rank number of the record in the other 

database if the pair is defined as a match. Zero is choosen as initial value. 

P_Selectivity and H_Selectivity indicate the selectivity between the best 

neighbour and the next best neighbour of a record. As initial value zero is 

chosen. 

P_Distfinal and H_Distfinal store the final distance of the matched pair. (It is not 

necessarily the distance beween the best neighbours). A very high distance is 

used as an initial value. 

If the pointer to the best neighbour indicates that no best neighbour was found 

in the first pass (P_Pointer1 = -1 or H_Pointer1 = -1) these records will be 

excluded from further analysis. (P_Lsi = 1; H_Lsi = 1); If at least a best 

neighbour was found the status for these records will remain as “undecided” 

(P_Lsi = -1; H_Lsi = -1) 

Determination of “Double zero” matches 

If a record has a distance of zero to its best neighbour and to its next best 

neighbour, it is uncertain which record has to be taken for the match. Both 

neighbours fit perfect. 

For this reason every record with such a double zero situation has to be marked 

as “double zeros” (.._LSI = 2). If the best neighbour of a record refers to a 

double zero match, this record also has to be marked with ..LSI = 2. 

This record points to a record in the other database pretending that this is a 

perfect fit, but the other record is highly uncertain. To put it dramatically, a 

double zero record is poisoning every record surrounding it. 

Records with status.._LSI = 2 have to be excluded from further processing as it 

is already known at this stage of the linking process that the situation will not 

change when running the second pass. Double zero records are a problem 

when linking the Austrian police database and the hospital database. The 

problem occurs due to lack of comparable information in both databases. 




Second pass 

In the second pass the determination of the records which should be linked is 

performed. This recursive procedure is almost the same as described in the 

paper of SWOV but enhanced with one function to find also matches between 

two records mentioned as next best neighbours in both databases. Figure 12 

shows a flow chart which describes this recursive algorithm. 

Figure 12: Description of the procedure used for the second pass 

Point 1: 

Find first record which is 

„undecided“ (_LSI = -1) 

Set Status = 1 

Point 2: 

Make this record the r.u.i. 

Status = 1 

Status 

Status = 2 

Look up r.u.i´s best 

neighbour 

Look up r.u.i´s next best 

neighbour 

yes 

_LSI of this 

neighbour is 

„undecided“ 

no 

yes 

_LSI of this 

neighbour is 


no 

yes 

First pointer 

points back to 

r.u.i. 

no 

Look up r.u.i´s next best 

neighbour 

Unmatchable: 

_LSI = “unmatchable” 

Go to Point 1 

Matched: 

_LSI = “matched” 

– go to Point 1 

Status = 1 

Go to Point 2 

yes 

__LSI of this 

neighbour is 


no 

yes 

Second pointer 

points back 

no 

yes 

First pointer 

points back 

no 

Matched: 


Go to Point 1 

Unmatchable: 


Go to Point 1 

Unmatchable: 


Go to Point 1 

Matched: 


– go to Point 1 

Status = 2 

Go to Point 2 

Description of names used: 

r.u.i 

Record under investigation 

Status 

status to detect if a pair who is pointing at each other as 

next best neighbour could be matched. 

First pointer H_Pointer1 or P_Pointer1 

Second pointer H_Pointer2 or P_Pointer2 

This procedure helps to find the following matches of records: 




1. if the record in the first database points to its best neighbour in the other 

database and this record also points back to the record in the first 

database as its best neighbour. 



database as its next best neighbour. 

3. if the record in the first database points to its next best neighbour in the 

other database and this record also points back to the record in the first 





Calculation of selectivity and final distance 

In this part of the linking procedure, the uniqueness of a match quantified by 

selectivity and final distance is calculated. First the difference between the 

distances of a record to its best neighbour and to its next best neighbour is 

calculated. This is done for records of both databases. The lowest difference of 

a matched pair is then stored in both records of the matched pair as their 

selectivity. To determine final distance the distance to the neighbour which is 

taken as matched neighbour is stored in each database. 

Taking into account selectivity and final distance, matched pairs can be divided 

in those which have a high probability to refer to the same person (Matchtype: 

“usable matched”) and those where it is doubtful if the match is ok (Matchtype: 

“not usable matched (Dist, Selectivity)”). A matched pair is “usable matched if 

the following conditions are fulfilled: 

• Final distance = 100 

These values have been chosen carefully after processing several tests. 

7.1.4 Results 

As no set of records with proofed quality of linking is available, the linking 

procedure could not be evaluated perfectly. To see if the linking procedure is 

calculating reasonable results, two alternative ways of checking the plausibility 

of the matched records is chosen. 

Reliability test I – Differences in affected body regions 

For this test we assumed that injuries of motorcyclists including moped riders 

and car occupants differ significantly when looking at the affected body region. 




Table 40 and Figure 13 show the mode of transport by affected body region for 

matched pairs with a matchtype “usable matched”. The standardised residuals 

show for example that knee and lower leg injuries of motorcyclists and moped 

riders differ significantly from the average due to the standardised residual of 

12,1. (standardised residuals +/- 1,94 would lead to a significance level of 5%). 

In all body regions rows the chi-square assumption is fulfilled. (Predicted value 

> 5) 

This leads to the assumption that the linking procedure does not match just 

records randomly. If so, there would be no significant difference by mode of 

transport. 

Table 40: Affected body regions of car passengers and motorcyclists 

(only usable matched persons: 3.624 records) 

Bodyregion * road user type Crosstabulation 

Bodyregion 

Total 

Abdomen injuries 

Ankel injuries 

Elbow injuries 

Hand injuries 

Head injuries 

Hip and thigh injuries 

Knee and lower leg 

injuries 

Neck injuries 

Shoulder injuries 

Thorax injuries 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

Std. Residual 

Count 

road user type 

Motorcyclist 

incl. Moped 

Car occupant rider Total 

239 91 330 

-,9 1,5 

48 40 88 

-2,4 4,3 

69 48 117 

-2,2 3,9 

51 32 83 

-1,6 2,8 

1167 210 1377 

3,5 -6,3 

100 59 159 

-2,0 3,6 

164 196 360 

-6,7 12,1 

299 12 311 

3,9 -7,1 

94 74 168 

-3,1 5,5 

543 88 631 

2,7 -4,9 

2774 850 3624 




Figure 13: Affected body regions of car passengers and motorcyclists 

incl. Moped riders (only usable matched persons: 3.624 records) 

Car occupants 

Motorcyclists incl. Moped riders 

Reliability test II – Linking different years 

For this test hospital data from the year 2001 were linked to police data from the 

year 2003. In the “date” - variable the year was changed from 2003 to 2001. 

Theoretically there should be no “useable matched” records after running the 

linking procedure. 

Table 41 shows that 1.567 records were found to be useable. This is 30,5% of 

the “useable matched” records when linking both databases with data from the 

year 2001. Therefore it can be assumed that 30,5% of these matched pairs do 

not refer to the same person in each database. This high share reveals that not 

enough information is available in both databases for getting reliable results. 




Table 41: Overview on linking procedure when linking police data from 

2003 with hospital data from 2001 

Source Matchtype Ergebnis 

Hospital double zeros 177 

not matchable 1258 

undecided 1789 

Hospital Ergebnis 3224 

Police Died on the scene 614 

double zeros 290 

not matchable 42065 

Police Ergebnis 42969 

Police + Hospital not usable matched (Dist, Selecticity) 13276 

usable matched 1567 

Police + Hospital Ergebnis 14843 

Gesamtergebnis 61036 

Overview and discussion 

Figure 14 and Figure 15 show the status of all records in the hospital and police 

database after the linking process. 

Figure 14: Status of records in the police database 

Status of records in the police database 

total: 57.223 

usable matched; 5142; 9% 

died on the scene; 685; 1% 

double zeros; 1425; 2% 

not usable matched (Dist, 

Selecticity); 9993; 17% 

not matchable; 39978; 71% 




Figure 15: Status of records in the police database 

Status of records in the hospital database 

total:18.067 

double zeros; 915; 5% 

usable matched; 5142; 28% 

not matchable; 2017; 11% 

not usable matched (Dist, 

Selecticity); 9993; 56% 

In the police database 71% of all records cannot be matched. As the hospital 

database contains only in-patients and no outpatients, this amount can be 

explained by the police coded “slightly injured” persons, who – in case they 

were coded properly – do not go to the hospital (73% of all records). 

It seems that we do have to less comparable information in both databases. 

This is indicated by the high percentage at 5% of all hospital records were the 

status is “double zero”. These records have exactly the same key variables in 

both databases. E.g. in Austria only the age, but not the date of birth is recorded 

in the police database. The linking procedure would be much more efficient if 

the police database would also contain this information. 

When looking at the total numbers, as presented in Figure 14 and Figure 15, it 

could be assumed that 915 records of double zeros from the hospital database 

refer to the same persons as 915 double zeros from the police database but, 

who fits with whom can not be determined. The remaining double zeros from 

the police database (1.425-915 = 510) can not treated as matchable. 

Preparation of the matrixes 

In order to prepare the matrixes a major problem is the lack of information about 

the mode of transport in the hospital database. E.g., it is known how many 

records in the police database refer to persons in a car. But there is no 

information on how many records in the hospital database refer to an accident 

where the person was sitting in a car. Therefore it is not possible to calculate 

how many people in a car were not coded by the police. Additionally, it is 




impossible to calculate the grey marked cells in the following table on the level 

of the mode of transport. 

Table 42: Possible combinations of presence or absence of hospitalised 

road traffic victims in police and hospital databases. (SWOV 2001) 

In police 

database 

Not in police 

database 

Not a 

hospitalised 

road traffic 

victim 

In hospital 

database 

In both 

databases 

Only in hospital 

database 

Hospitalisation 

not caused by 

a road traffic 

accident 

Not in 

hospital 

database 

Only in 

police 

database 

In neither 

database 

Not a 

hospitalised road 

traffic victim 

Road traffic 

victim but not 

hospitalised 

In order to provide a full picture of the linking procedure every record of both 

databases is counted in the matrixes. If two records are matched “usable 

matched” or “not usable matched” all needed cells of the matrixes can be 

calculated. In all other cases there is a lack of information and cells, which could 

not be calculated, are marked with “n/a” 

Additional information to the variables proposed by TRL concerning the design 

of matrix 1 and matrix 2 is provided. These are: 

- Variable Source: This indicates wether this is a matched pair (Police + 

Hospital) or a record could not find another record (Police; Hospital) 

- Variable Matchtype: 

• Died on the scene 

• Double zeros 

• Not matchable 

• Undecided 

• Not usable matched (Dist, Selectivity) 

• Usable matched 

This information is needed to analyse why data are not matched properly. 




Matrix 1 for the matched pairs 

Table 43 shows an overview on matrix1 where police coding, length of stay in 

hospital and road user type can be crossed. 

Table 43: Matrix 1 overview 

Source Matchtype police coding Length of stay SafetyNet Ergebnis 

Hospital double zeros n/a > 3 days 370 

1 - 3 days 310 

Outpatients; treated like inpatients 81 

Overnight 154 

not matchable n/a > 3 days 322 

1 - 3 days 295 


Overnight 154 

undecided n/a > 3 days 466 

1 - 3 days 431 


Overnight 195 

Police Died on the scene Fatal Injured n/a 685 

double zeros Fatal Injured n/a 16 

Injury unknown (seriously injured) n/a 316 

Seriously Injured n/a 550 

Slightly Injured n/a 543 

not matchable Fatal Injured n/a 132 

Injury unknown (seriously injured) n/a 3608 

Seriously Injured n/a 3889 

Slightly Injured n/a 32349 

Police + Hospital not usable matched (Dist, Selecticity) Fatal Injured > 3 days 12 

1 - 3 days 12 


Overnight 5 

Injury unknown (seriously injured) > 3 days 486 

1 - 3 days 536 


Overnight 359 

Seriously Injured > 3 days 585 

1 - 3 days 427 


Overnight 240 

Slightly Injured > 3 days 2309 

1 - 3 days 2751 


Overnight 1668 

usable matched Fatal Injured > 3 days 19 

1 - 3 days 13 


Overnight 12 

Injury unknown (seriously injured) > 3 days 361 

1 - 3 days 479 


Overnight 292 

Seriously Injured > 3 days 1447 

1 - 3 days 699 


Overnight 244 

Slightly Injured > 3 days 288 

1 - 3 days 583 


Total 

Overnight 487 

60155 

If a person is released from the hospital within the same day, one day is stored 

in the variable “XDauer” of the Austrian hospital database. If a person is 

overnight in the hospital two days are stored in this variable. The conversion 

from “Xdauer” to the definition of “Length of stay SafetyNet” used in this report is 

as follows: 




Table 44: Transformation of “Xdauer” to “Length of stay SafetyNet” used 

in this report 

Xdauer Length of stay SafetyNet 

Outpatients; treated like 

1 

inpatients 

2 Overnight 

3 1 - 3 days 

4 1 - 3 days 

5 1 - 3 days 

6 and more > 3 days 

1.434 persons are coded “slightly injured” in the police database but these 

persons went to the hospital after the accident. This shows that the police 

underestimated the severity of injuries. 

In the police database 132 persons coded “fatal injured” are not matchable at 

all. This could be a topic for further investigations as the fatalities stored in the 

police database are often cross-checked with the Austrian mortality statistic. As 

this statistic is used for the population statistic, it can be assumed that the 

quality of this database is very high. Therefore, the high share of 132 missing 

persons in the hospital database is doubtful. E.g. further investigations on a 

single record revealed that this person was transfered to hospital A after the 

accident but transfered to hospital B after some days. In contradiction to the 

description of the Austrian hospital discharge database there was no record 

about the stay of the person in hospital A - only about hospital B. 

Figure 16 shows the distribution of length of stay in hospital by mode of 

transport. A higher proportion of of a long stay in hospitals can be shown for 

pedestrians and motorcyclists including moped riders. 

Figure 16: Distribution of length of stay in hospital by mode of transport 

(only usable matched persons: 5.142 records) 

60,0% 

54,3% 

52,2% 

50,0% 

45,3% 

40,0% 

30,0% 

20,0% 

13,8% 

28,0% 

33,9% 

17,8% 

29,4% 

15,7% 

23,1% 

37,2% 

34,8% 

36,8% 

38,8% 

18,8% 

Outpatients; treated like inpatients 

Overnight 

1 - 3 days 

> 3 days 

10,0% 

3,9% 3,0% 2,8% 

4,8% 5,6% 

0,0% 

Pedestrian Pedal cyclist Motorcyclist incl. 

Moped rider 

Mode of transport 

Car occupant 

Other 




Matrix 2 for the matched pairs 

Conversion of ICD-10 Codes to MAIS 

In Austria the international statistical classification ICD-10 is used in the hospital 

discharge database. Before 2001 hospitals reported the main diagnosis by 

using the classification ICD-9. Therefore, a transformation from ICD-10 to MAIS 

is carried out. 

Within the EC project Apollo a conversion from ICD-10 diagnosis to ISS has 

been developed at the University of Navarra. Based on this code a conversion 

from ICD-10 to MAIS is used for the Austrian hospital discharge diagnosis. Not 

for all ICD-10 codes MAIS can be calculated. In such cases the value of MAIS is 

“n/a” 

Problems occur when adapting the Apollo code to this data. The Apollo code 

was developed to use up to three diagnoses per record. In the Austrian hospital 

discharge database only the main diagnosis is stored. It could be assumed that 

this is the reason why the conversion does not work very properly. 

To show the problem, records of the hospital database are chosen, where there 

is an indication in the hospital database that this person left the hospital due to 

death. In the database 171 records are marked with this value. Out of these 

records 126 records show a diagnosis which can be transformed to MAIS. The 

following table shows the number of cases and the calculated MAIS values. For 

only one record, a MAIS score of 6 can be calculated. As the source for this 

table is only the hospital database all records should have a MAIS value of 6 as 

these persons left the hospital dead. 

Table 45: MAIS scores of persons who died in the hospital 

MAIS Score Number of records 

1 10 

2 40 

3 45 

4 12 

5 18 

6 1 

Total 126 

Therefore, the conversion of the ICD-10 codes to MAIS using only one 

diagnosis, as in the Austrian hospital database, could be criticised. 




Calculated Matrix 2 

Table 46 shows a matrix where mode of transport, police coding and MAIS is 

crossed. Only “useable matched” pairs are used in this table. 

Table 46: Overview Matrix 2 (only useable matched records) 

Sum of Counts 

MAIS 

road user type police coding 1 2 3 4 5 6 n/a Total 

Car occupant Fatal Injured 5 17 11 4 2 14 53 

Injury unknown 357 305 34 1 1 58 756 

Seriously Injured 273 700 135 6 16 57 1187 

Slightly Injured 497 361 19 2 3 73 955 

Motorcyclist incl. Moped rider Fatal Injured 2 3 1 1 7 

Injury unknown 70 95 10 1 8 184 


Slightly Injured 71 46 6 14 137 

Other Fatal Injured 1 1 2 

Injury unknown 19 27 3 1 7 57 

Seriously Injured 36 72 14 2 9 133 

Slightly Injured 37 15 1 5 58 

Pedal cyclist Fatal Injured 1 5 2 8 



Slightly Injured 58 70 8 1 1 13 151 

Pedestrian Fatal Injured 1 6 5 2 2 2 18 



Slightly Injured 53 61 3 1 15 133 

Total 1728 2579 415 35 54 1 330 5142 

Figure 17 presents the distribution of MAIS by different modes of transport. 

Although the transformation of ICD-10 to MAIS is doubtful, differences can be 

observed. Car occupants have a higher share of MAIS 1, especially compared 

to pedestrians. 

Figure 17: Distribution of MAIS by mode of transport (only usable matched 

persons: 5.142 records) 

70,0% 

60,0% 

50,0% 

40,0% 

30,0% 

20,0% 

10,0% 

0,0% 

55,6% 

52,7% 

57,5% 

29,1% 

25,6% 

23,7% 

11,6% 

10,6% 

9,2% 

7,2% 

6,7% 

6,0% 

8,8% 

5,8% 6,8% 

4,2% 

1,5% 1,7% 0,0% 

2,0% 

1,4% 1,3% 0,7% 1,6% 

0,0% 

0,4% 0,1% 

0,4% 0,0% 

0,0% 0,0% 

Pedestrian Pedal cyclist Motorcyclist incl. 

Moped rider 


38,4% 

46,9% 

Car occupant 

36,8% 

45,6% 

Other 

MAIS 1 

MAIS 2 

MAIS 3 

MAIS 4 

MAIS 5 

MAIS 6 

MAIS n/a 




7.1.5 Conclusions 

The computer assisted linking process of the Austrian hospital and the police 

database described in this report tried to calculate two matrixes which should 

enable further calculation of underreporting rates (conversion factors). 

The design of the calculated matrixes fits the requirements of SafetyNet WP1 

Task 5. Nevertheless, the calculated output has to be treated with care. 

“Useable matched” calculated records are likely not always to refer to the same 

persons. It can be assumed that 30,5% of these matches are wrong. The 

reason for this is not the linking procedure itself, but the lack of comparable 

information on both databases. This lack explains the high share of “not useable 

matched” records due to selectivity, distance (56%) and the “double zero” 

records (5%). 

The calculated MAIS is based on a transformation of ICD-10 Codes to MAIS. 

This transformation produces doubtful results. Therefore, no conversion factors 

based on MAIS can be derived from this linking process. 

To calculate underreporting rates for different modes of transport, information 

about the mode of transport is needed in both databases. The Austrian hospital 

discharge database does not contain this information. No conversion factors for 

different modes of transport can be calculated with the result of this report. As 

the Austrian hospital discharge database contains only in-patients, information 

about slightly injured persons (who do not go to the hospital) are not treated. 

Future steps & recommendations: 

- Results of the linking procedure could be highly improved if e.g. the date of 

birth is also mentioned in the police database. A project called “UDM – 

Unfalldatenmanagement” with the aim of changing the paper based accident 

investigation into a computer based investigation could change this situation. 

- More analysis is required to detect why there are so many “unmatchable” 

records. 

- More investigation about the recording procedure of persons, who are 

admitted to a hospital, but than transferred to another hospital later, is 

needed. 

- The new IDB (Injury Database) of DG-Sanco will also contain information 

about the mode of transport in traffic accidents in near future. Unlike the 

Austrian hospital discharge database, the IDB is based on a survey carried 

out in different member states of the EU. More information about the 

accident will be available with this new database which could be used to 

relaunch of the calculation of underreporting rates. 




7.1.6 References 

Apollo Project: http://www.unav.es/preventiva/traffic_accidents/pagina_5.html 

University of Navarra – Apollo Project, visited on 5 th of January 2007. 

Clark, DE. (2004). Practical introduction to record linkage for injury research. 

Injury Prevention 2004, Vol. 10, pp186-191. 

Crash Outcome Data Evaluation System (CODES): http://wwwnrd.nhtsa.dot.gov/departments/nrd-30/ncsa/codes.html, 

NHTSA, visited on 20 th 

of June 2007 

Howe GR. Use of computerized record linkage in cohort studies. Epidemiol Rev. 

1998, 20(1):112-21. 

Jaro MA. Probabilistic linkage of large public health data files. Stat Med. Mar 15- 

Apr 15 1995, 14(5-7):491-8. 

SWOV (2001). A new linking procedure for the determination of the total 

number of hospitalised road traffic victims by comparing police and hospital 

reports. Netherlands 




7.2 Study carried out in Czech Republic 

Report prepared by Jan Tecl (CDV) 


The Czech national study comparison and linking the records of road traffic 

victims from the police database and the records of treated or hospitalised 

persons from the hospital database was carried out in the framework of 

SafetyNet project as probably the first study of this kind in the Czech Republic in 

2007. 

The two datasets with details of injured persons were for the Czech district of 

Kromeriz in the years 2003 – 2005. Records from the central police database 

and the district hospital were used. 


The police database 

The police database of the Czech Republic is widely believed to be a quite 

reliable road accident data source, and has been developed over more than 30 

years. It is organised in three levels: district, regional and central. Data collected 

at the district level are verified and transmitted to the regional level, again 

verified and transmitted to the central level, again verified and then stored to the 

central database on the State Police Directory. The data collecting process is 

carried out by means of the road accident form and this process is now highly 

computerised. The accident report form contains 74 variables related to the 4 

main groups (accident, vehicle, driver / passenger, pedestrian). 

By law, there is a legal obligation to report to the police all accidents on public 

roads, not only those with any person injured but also those with only material 

damage (with damage over defined financial limit - at present 50 000 CKR, 20 

000 CKR from 2001 until 2006 and only 1000 CKR before 2001. 

The hospital discharge database 

The situation for the hospital data for road accident victims is not so clear in the 

Czech Republic. There is no central hospital injury database as in other 

countries. Nevertheless, the statistics of hospitalised and killed persons from all 

hospitals in the country are collected by the Institute of Health Information and 

Statistics, that is the administrator of the National Health Information System. 

The following reports are, among others, processed by this Institute: 

Statistics of deceased (national demographic statistics data of deceased 

persons, amended by the cause of death by diagnosis in ICD-10 system) and 

National register of hospitalised (individual hospitals data about hospitalised 

persons, basic diagnosis and possibly also other diagnoses, possibly operation 

and death cause diagnoses). 




The problem is that only data are available from this Institute (summary 

statistical publications), but disaggregate data records are available - even 

without personal identifiers. This strict personal data protection is defined by the 

law. 

From the local point of view, the statistics of separate hospitals vary 

considerably in their structure and reliability, so it is very difficult to obtain from 

here some reasonable and utilizable data. Finally, one hospital from Kromeriz in 

central Moravia (about 70 km from Brno) was chosen, as the example of the 

hospital with relatively reliable and available data. This hospital operates for the 

town with 30 000 inhabitants (and the neighbouring area), which is about 0,3 % 

of total inhabitants of the Czech Republic. 

There are two main problems. Firstly, it is uncertain how closely the operational 

zones of the police district and the hospital match. Secondly, the hospital 

statistics are predominantly oriented to the medical elements of the case 

(diagnosis of the injury), while the elements related to the accident 

circumstances are not the focus of attention of the medical personnel, even in 

the hospitals with better level of data collection. 


Variables used for the linking process 

The best linking variable would be surely the ID-number of injured person. This 

possibility is, of course, excluded by the law for personal data protection. 

Consequently, alternate variables must be used. 

Although most of the variables in the police and hospital databases are disjoint, 

some corresponding variables could be found. The following variables have 

been chosen for the linking process: 

• date of accident 

• year of birth, 

• sex, 

• type of road user. 

Skateboard and roller skate users are registered in the hospital database as a 

specific type of road user, but they are considered as pedestrians for the linking 

procedure. 

Some tolerance is allowed for variables in the linking process: 

day of accident: +1 day in the hospital database, 

year of birth: ± 5 years 

type of road user: some difference in the hospital database is allowed 




Other important variables which are necessary for the final evaluation and 

comparison, are the injury severity (from the police database) and the length of 

stay in the hospital and MAIS (Maximum Abbreviated Injury Scale). MAIS is 

derived from the ICD-10 system code. The codes used are: 

Road user type: 

1=pedestrian 

2=pedal cyclist 

3=motorcyclist 

4=car occupant 

9=other 

Police severity: 

1=fatality 

3=seriously injured 

4= slightly injured 

-1= not matched 

MAIS: 

1=MAIS 1 

2=MAIS 2 

3=MAIS 3 

4=MAIS 4 

5=MAIS 5 

6=MAIS 6 

9=unknown MAIS' 

-1=not matched 

Length of hospital stay: 

1='outpatient' (0 night) 

2='overnight' (1 night) 

3='hospitalised 1-3 days' (1-3 nights) 

4='hospitalised more than 3 days' (>4 nights) 

8=’hospitalised but unknown length' 

-1='not matched' 

The third party parameter has not been taken into consideration because there 

is no corresponding value in the hospital database. 

The linking procedure 

The procedure used for the data linking is a probabilistic method accomplished 

by the semi-automatic way with a manual checking of linked records. The data 

of both groups (police and hospital) are ordered by the date of the accident. 

Then the records with the same or near linking parameters are gradually offered 

for the linking in two passes. The definitive linkage can be accepted or rejected. 




7.2.4 Results 

Table 47: Total number of linked records 

police 

data 

no police 

data total 

hospital data 266 575 841 

no hospital data 845 - 

total 1111 1686 

matched records 

In total, 266 from 1111 (23,9%) police records were matched among the 

hospital records. 266 from 841 (31,6%) hospital records were matched among 

the police records. The total registration rate is 65,9% (1111/1686). 

Table 48: Number of linked records for type of user 

police and only police only hospital police hospital 

type of user hospital data data data total total total 

pedestrians 28 68 62 96 90 158 

pedal cyclists 60 135 422 195 482 617 

motorcyclists 19 74 15 93 34 108 

car occupants 148 523 76 671 224 747 

other 11 45 0 56 11 56 

total 266 845 575 1111 841 1686 

The registration rate is 60,8% for pedestrians, 31,6% for pedal cyclists, 86,1% 

for motorcyclists, 89,8% for car occupants and practically 100% for other road 

users. 




Table 49: Number of linked records for age and sex 

age / sex 

police and 

hospital data 

only police 

data 

only hospital 

data police total hospital total total 

men 

0-14 10 37 62 47 72 109 

15-24 44 173 90 217 134 307 

25-49 71 258 140 329 211 469 

50-69 30 76 63 106 93 169 

70+ 12 28 13 40 25 53 

women 

0-14 

15-24 

7 

26 

21 

67 

37 

43 

28 

93 

44 

69 

65 

136 

25-49 34 105 76 139 110 215 

50-69 27 57 38 84 65 122 

70+ 5 23 13 28 18 41 

total 266 845 575 1111 841 1686 

The registration rate is 43,1% for men aged 0-14, 70,7% for men aged 15-24, 

70,1% for men aged 25-49, 62,7% for men aged 50-69 and 75,5% for men at 

least 70 years old (66,8% for all men). The registration rate is 43,1 % for women 

aged 0-14, 68,4% for women aged 15-24, 64,7% for women aged 25-49, 68,9% 

for women aged 50-69 and 68,3% for women at least 70 years old (64,2% for all 

women). 




Table 50: Matrix for MAIS and police severity linking 

police severity 

MAIS 1: fatal 3: serious 4: slight -1: not matched 

1: pedestrians 

1 

2 

0 

0 

1 

2 

9 

8 

46 

9 

3 0 4 1 6 

4 0 0 0 0 

5 1 1 0 0 

6 1 0 0 1 

9 0 0 0 0 

-1 4 20 44 0 

2: pedal cyclists 

1 

2 

0 

0 

0 

3 

35 

12 

358 

52 

3 0 3 2 10 

4 0 3 0 2 

5 0 1 0 0 

6 1 0 0 0 

9 0 0 0 0 

-1 5 20 110 0 

3: motorcyclists 

1 

2 

0 

0 

2 

0 

8 

5 

13 

2 

3 0 1 1 0 

4 0 0 0 0 

5 0 1 0 0 

6 1 0 0 0 

9 0 0 0 0 

-1 2 18 54 0 

4: car occupants 

1 

2 

0 

0 

4 

7 

84 

37 

61 

12 

3 0 5 4 2 

4 0 1 1 0 

5 2 1 0 1 

6 2 0 0 0 

9 0 0 0 0 

-1 18 64 441 0 

9: other 

1 

2 

0 

0 

0 

0 

6 

4 

0 

0 

3 0 1 0 0 

4 0 0 0 0 

5 0 0 0 0 

6 0 0 0 0 

9 0 0 0 0 

-1 0 7 38 0 




Table 51: Matrix for length of stay and police severity linking 

police severity 

length of stay 1: fatal 3: serious 4: slight -1: not matched 

1: pedestrians 

1: outpatient 

2: overnight 

1 

0 

1 

0 

11 

1 

52 

1 

3: 1 - 3 days 0 1 5 6 

4: >3 days 1 6 1 3 

8: unknown 0 0 0 0 

-1: not matched 4 20 44 0 

2: pedal cyclists 

1: outpatient 

2: overnight 

1 

0 

2 

0 

40 

0 

386 

2 

3: 1 - 3 days 0 3 8 22 

4: >3 days 0 5 1 12 

8: unknown 0 0 0 0 

-1: not matched 5 20 110 0 

3: motorcyclists 

1: outpatient 

2: overnight 

1 

0 

2 

0 

11 

0 

14 

0 

3: 1 - 3 days 0 0 2 1 

4: >3 days 0 2 1 0 

8: unknown 0 0 0 0 

-1: not matched 2 18 54 0 

4: car occupants 

1: outpatient 

2: overnight 

2 

0 

5 

0 

93 

0 

67 

0 

3: 1 - 3 days 0 6 27 6 

4: >3 days 2 7 6 3 

8: unknown 0 0 0 0 

-1: not matched 18 64 441 0 

9: other 

1: outpatient 

2: overnight 

0 

0 

0 

0 

9 

0 

0 

0 

3: 1 - 3 days 0 1 1 0 

4: >3 days 0 0 0 0 

8: unknown 0 0 0 0 

-1: not matched 0 7 38 0 


The linking process has been carried out for the first time in the Czech Republic. 

It seems however, that police data are significantly more reliable than hospital 

data about accidents victims because the hospital data are not collected so 

carefully. It would be necessary, for more accurate results, to improve and unify 

the system of collecting hospital accident statistics. Further, the linking 




procedure may have been affected by the possibility that the not catchment 

area may not correspond fully. 

It can be seen from the results that the lowest registration rate is for bicyclists 

(probably most of them were injured in accidents involving a single bicycle) and 

then for pedestrians (but sometimes it not clear from the hospital database if the 

pedestrian was really injured in a traffic accident that involved another person). 




7.3 Study carried out in France 

Report prepared by Emmanuelle Amoros (INRETS) 


WP1 Task 1.5 aims at estimating factors to correct for under-reporting, to be 

applied on the CARE data. As in the other countries, the CARE data for France 

are made of the national police data. The medical data necessary for 

comparison with the police data are provided in France by a road trauma 

registry. This registry covers all victims of road crashes that occurred in the 

Rhône county and who seek medical attention in health facilities of the county 

or close surroundings. The Rhône county is a large county of 1.6 million 

inhabitants. It consists of a large city Lyon, its suburbs and a rural area in the 

north part. 

The record-linkage of the police data and the registry data is being conducted 

every year, as a routine. It is described below. Results of the linkage are also 

provided. 

Only the French specificities of the data and of the record-linkage are described 

here. The common methodology used to estimate the under-reporting correction 

factors is described elsewhere. 


Police data 

The French police are required by law to write a crash report for every road 

crash causing at least one casualty. A road crash is officially defined as a crash 

involving at least one vehicle and occurring on the network open to public traffic. 

Skateboard or roller skate users are considered as pedestrians by the police, 

and, as such, are only classified as road casualties if hit by a vehicle. There is 

no restriction about motorised vehicles, in other words there is no exclusion 

criteria on bicycles. 

The police crash report should report all the people involved in the crash: killed, 

injured and non-injured ones. Injured are classified into slightly or seriously 

injured : casualties requiring a hospital stay of 6 days or more are categorised 

as ‘seriously injured’, whereas casualties requiring less than 6 days of hospital 

stay (including outpatients) are categorised as ‘slightly injured’. 

The police crash report contains detailed information on the crash, the crash 

environment and conditions, the vehicles involved, but it contains limited 

information on the people involved. 

These police reports are paper reports; most of the information they contain is 

recorded into electronic files, according to a standardised format. 

The police dataset used here is the one restricted to the Rhône county: only 

crashes that occur in the Rhône county are selected. 




Hospital data (road trauma registry) 

The registry covers all casualties from road crashes in the Rhône county who 

seek medical attention in health facilities. Inclusion criteria are broader than the 

police ones: off-road crashes are not excluded; roller skate, skate-board or 

scooter users are not considered as pedestrians but as road users using a 

mean of transport and are hence included, whether hit by another vehicle or not. 

The registry is based on the participation of all health care facilities in the county 

(and its close surroundings) that may receive victims of a traffic crash: it 

includes some 150 health care facilities: from emergency departments, 

intensive care units, surgery units... to rehabilitation departments, as well as 

pre-hospital emergency care. The registry includes both inpatients and 

outpatients, i.e. all casualties, whether hospitalised or not. 

Information collected for each casualty consists of a few crash characteristics 

and of the following casualty characteristics: gender, date of birth, place of 

residence, hospital stay, hospital transfer if relevant, and accurate injury 

assessment. Indeed, for each subject, injury assessment is based on the whole 

set of diagnoses provided by the different health services the subject may have 

gone through. Plain text diagnoses are coded by the registry physicians 

according to the Abbreviated Injury Scale (AIS), 1990 revision. Each injury is 

assigned a severity score, ranging from AIS 1 (minor) to AIS 6 (beyond 

treatment). To measure the overall severity for casualties with multiple injuries, 

the MAIS is used (maximum AIS severity score). 

Length of hospital stay is estimated by the number of nights spent at the 

hospital, with the number of nights being obtained from date of discharge and 

date of admission. 

The categories used are the following: 

- outpatient (0 night) 

- overnight (1 night) 

- 1-3 days (2-4 nights) 

- >3 days (>4 nights) 

- inpatient, but unknown length of stay 

Preparation of the data 

The data cover the 1996-2003 time period. The (French) police definition of 

road casualties are applied, as these are the definition of the data in the CARE 

database. It implies that the following road users have been excluded: 

- roller, scooter and skate boards users if not hit by a vehicle, 

- uninjured road users. 

Fatalities have also been excluded since they have high rates of reporting. 





Methodology 

We have first implemented a record-linkage methodology (Clark, 2004) to be 

applied retrospectively on the police and medical datasets, once they are 

available. The 1996-2001 data have been linked this way. The method used is 

both probabilistic and manual. 

It is manual because a major linking variable (place of accident: city/village and 

details such as street name) cannot be standardised into numerical codes. This 

variable is unformatted free text which cannot be standardised and coded 

without important loss of information. 

The method is manual in the sense that the person in charge of the linking 

process goes through any single record (of the police dataset), trying to match it 

with one from the medical registry dataset; this process is performed on the 

computer screen, assisted by a user-friendly specific application. 

It is a probabilistic method as we allow for some possible error in the linking 

variables. Typically, we allow for the date of accident to differ by 1 or 2 days, if 

the other linking variables agree… No matching weights based on probabilities 

are computed (since it was not possible on one of the linking variables). The 

decision of linking two records is made based on how many linking variables 

agree, which ones and on which values. 

The linking variables are : 

• date of crash, 

• time of crash, 

• location of crash (town/district/village and details such as road(s) number 

or street(s) name), 

• date of birth (only year and month are available) of the casualties, 

• gender of the casualties, 

• road user type of the casualties 

the most important ones being date of crash, location of crash, year and month 

of birth of casualties 

From 2002 onwards : 

In order to improve the exhaustiveness of the registry and the completeness of 

its information (i.e. reduce the number of missing values), the use of the police 

data is now part of the registry recording procedure. That is to say, every time a 

casualty is about to be recorded in the registry (from a notification form), it is 

first checked whether this casualty can be found in the police dataset. If so, the 

registry record is created by specifying the link with the record found in the 

police data and by copying police information about the crash (location, type of 

crash). If the casualty is not found in the police data, the registry record is 

created, using information from the notification form. 




Linking software 

A specific software was developed in Visual Basic ; it works in the Microsoft 

Access environment. 

The software is basically a user-friendly way of comparing the two datasets. It 

allows for different sorting on the linking variables, and pre-selecting of the 

records that match on date of accident and date of birth (year and month). It 

displays casualties records grouped within accident. Values of all different 

linking variables are displayed. One goes through every police casualty record, 

tries to find the corresponding record in the medical registry dataset. Two 

records are linked with a “press-button”, and hence be selected out of the 

records to be linked. 

From 2002 onwards : 

A specific software was developed in Visual Basic ; it works in the Microsoft 

Access environment. It works very much in the same way. 

Problems encountered and solutions 

Crash location is missing in 16% of the medical registry casualties. Since crash 

location is a major linking variable, a missing value often means that this 

casualty will not be linked. One must however keep in mind that at best, since 

the police file is not even half the size of the registry file, one could match about 

one registry casualty out of two. In other words, 16% of missing crash location 

(in the registry casualties) does not mean 16% of missed links. This was 

confirmed (see section 5 – reliability of the linkage) 

There is nothing much that can be done about this missing crash location. The 

registry staff already sends letters to the casualties when some data is missing, 

and obtains some but not all. 

7.3.4 Results 

The results of the linkage are provided in the following tables: 

Table 52: Number of linked records 

Number of records from 

police data 

Number of records from 

hospital data 

Linked 21 310 (62.7%) 21 310 (27.0%) 

Non-linked 12 668 (37.3%) 56 479 (73.0%) 

Total 33 978 (100 %) 77 789 (100 %) 

In the police data (restricted to injured), a little less than two thirds were linked 

with the hospital data. Conversely, less than one third of hospital data were 

linked with hospital data. 




Table 53: Police data and link status according to police severity 

Police reported 

severity 

Number 

of records 

Proportion 

linked 

Killed (at 6 days) excluded - 

Seriously injured 4 862 75.4% 

Slightly injured 29 116 60.6% 

Non-injured excluded - 

Total 33 978 62.7% 

The proportion of police records linked to the hospital data increase with police 

severity classification. 

Table 54: Hospital data and link status according to MAIS 

Length of stay 

Number of 

records 

Proportion 

linked 

MAIS 1 56 043 22.9 

MAIS 2 15 535 34.9 

MAIS 3 4 110 53.1 

MAIS 4 752 64.6 

MAIS 5 and 6 248 71.4 

Unknown MAIS 1 100 16.8 

total 77 789 27.0 

The proportion of hospital records linked to police records increases with the 

MAIS, from 23% at MAIS 1 to 71% at MAIS 5-6. 

Table 55: Hospital data and link status according to length of hospital stay 



records 

Proportion 

linked 

Killed (at 6 days) excluded - 

outpatients 64 112 23.6 

overnight 5 062 36.9 

hospitalised 1-3 days 2 799 43.0 

hospitalised > 3 days 4 187 59.8 

Hospitalised, unknown length of stay 1 627 33.9 

total 77 789 27.0 

Similarly, the proportion of hospital records linked with the police records 

increases with length of hospital stay. 




Casualties characteristics according to registration 

We provide distributions of a number of casualties characteristics, according to 

source registration. This is defined in 3 groups: police-only data, hospital-only 

data and intersection between police and hospital data. Being based on 

independent data the 3 groups can be compared. The comparison of casualties 

characteristics between hospital-only and ‘police intersection hospital’ provides 

a description of the police “filter” on casualties and hence enables the 

identification of bias factors for police under-reporting. 

Table 56: Road user type according to registration source 

Road user type Police only Police ∩ hospital Hospital only 

pedestrians 13.3 14.5 7.7 % 

cyclists 3.0 3.5 17.7 % 

motorised-two 

15.9 17.8 20.7 % 

wheelers 

car occupants 63.4 60.6 49.9 % 

others 4.4 3.5 4.0 % 

100.0 % 

(n=12 668) 

100.0 % 

(n= 21 310) 

100.0 % 

(n= 56 479) 

There is hardly any difference in road user type distribution between police-only 

casualties and police∩hospital casualties. This indicates an absence of bias on 

road user type in hospital reporting. 

On the contrary, the distributions of road user type are different between 

hospital-only and police∩hospital : there are far fewer cyclist casualties in the 

police intersection hospital data than in the hospital only data. This indicates a 

bias towards less reporting of cyclists compared to other road user types in the 

police data. 

Table 57: Presence of third party, according to registration source 

Third party Police only Police ∩ hospital Hospital only 

yes 84.2 84.1 55.7 % 

no 15.8 15.9 44.3 % 

total 100.0 % 

(n=12 668) 

100.0 % 

(n= 21 310) 

100.0 % 

(n= 56 479) 

What we mean by ‘third party’ is whether a (human) opponent no matter is 

he/she was injured or not, and no matter if he/she was a pedestrian, or 

someone in a vehicle, whether motorised or not, and whatever the type (bicycle, 

car, van , bus, train, tram..). 




There is no difference in third party distribution between police-only casualties 

and police ∩ hospital casualties, indicating no bias on the existence of third 

party in the crash in hospital reporting. On the contrary, these third party 

distributions are highly different between the police∩hospital data and hospitalonly 

data: a much higher proportion of casualties with no third party involved in 

the hospital-only data compared to the police data. This indicates a bias 

towards lower reporting in the police data of casualties involved in crashes with 

no third party than casualties involved in a crash with a third party. 

Table 58: Crash location: road type, according to registration source 

Road type Police only Police ∩ 

hospital 

Hospital 

only 

Hospital only 

motorways 12.2 11.1 6.7 9.8 

state and 

30.0 31.9 7.6 11.2 

county roads 

local roads 54.6 54.2 49.7 72.8 

off-road, other 3.2 2.7 4.2 6.2 

unknown 0.0 0.0 31.7 Not accounted for 

100.0 % 

(n=12 668) 

100.0 % 

(n= 21 310) 

100.0 % 

(n= 56 479) 

100.0% 

(n= 38 554) 

There is hardly any difference in road type distribution between police-only 

casualties and police ∩ hospital casualties, indicating no bias on road type 

(where the crash occurred) in hospital reporting. 

The further comparison of road type distribution is hindered by a large 

proportion of unknown in the hospital-only data. However it seems that the 

distribution of road type is different between hospital-only data and police 

intersection hospital data: there would a smaller proportion of casualties from 

crashes on local roads and off-road in the police data than in the hospital-only 

data. This indicates that the police under-reporting is worse for those crashes. 

In other words, there is some bias on road type in police reporting. 




Table 59: Crash location: “greater Lyon: inside or outside” according to 

registration source 

Urban/rural Police only Police ∩ 

hospital 

Hospital 

only 

Hospital only 

Lyon 34.8 32.6 19.5 24.7 

Lyon suburbs 39.1 47.3 35.6 45.1 

outside ‘greater 25.9 19.9 23.8 30.1 

Lyon’ 


100.0 % 

(n=12 668) 

100.0 % 

(n= 21 310) 

100.0 % 

(n= 56 479) 

100.0% 

(n= 44 557) 

There is some difference in the distribution of where the crash occurred 

between police-only casualties and police ∩ hospital casualties: casualties 

involved in crashes that occurred far from Lyon are less often found in hospital 

data: it is probable that slight casualties who crashed far from a hospital are less 

likely to go to the hospital than similarly slight casualties who crashed within 

‘greater Lyon’. In other words there is a bias on urban/rural area in hospital 

reporting. 

Again, there is a large proportion of unknown in the hospital-only data. However 

the distributions of casualties according to “greater Lyon” (crash location) seems 

different between hospital-only and Police ∩ hospital: the proportion of 

casualties who crashed outside the “greater Lyon” is smaller in the police ∩ 

hospital than in the hospital only. This indicates a worse police under-reporting 

for these casualties. More generally, it means a bias on this characteristic in 

police-reporting too. 

Table 60: Crash location: police type area, according to registration 

source 

Police type 

area 

urban police 

area 

rural police 

area 

urban 

motorway 

police area 

Police only 

Police ∩ 

hospital 

Hospital 

only 

Hospital only 

61.0 61.4 36.1 49.8 

22.5 21.5 30.3 41.9 

16.5 17.0 6.0 8.2 


100.0 % 

(n=12 668) 

100.0 % 

(n= 21 310) 

100.0 % 

(n= 56 479) 

100.0% 

(n=40 915) 




There is no difference in police type distribution between the police-only 

casualties and the police intersection hospital casualties: this indicates that 

there is no bias on police type area in hospital reporting. 

Again, there is a large proportion of missing data in hospital-only data. However 

the distributions of police type seem different. There is a higher proportion of 

casualties from rural police area in the registry than in the police intersection 

hospital casualties: this indicates a bias on police type in the reporting of 

casualties in the police data. More precisely, the under-reporting is worse in 

rural police area. This corroborates with the previous finding of worse underreporting 

outside the “greater Lyon” area. 

Table 61: Age of casualty, according to registration source 

Age Police only Police ∩ hospital Hospital only 

0-14 6.5 7.5 13.4% 

15-24 29.6 30.7 30.3% 

25-49 43.8 43.7 38.2% 

50-69 13.4 12.6 8.8% 

70 and over 5.6 5.1 3.0% 

unknown 1.1 0.4 2.8% 

100.0 % 

(n=12 668) 

100.0 % 

(n= 21 310) 

100.0 % 

(n= 56 479) 

There is hardly any difference in age distribution between police-only casualties 

and police ∩ hospital casualties. This indicates no bias on age in hospital 

reporting. There is some difference in age distribution between hospital-only 

casualties and police ∩ hospital casualties: younger casualties in the hospitalonly 

data. This indicates a slight bias on age in police reporting. 

Table 62: Gender of casualty according to registration source 

Gender Police only Police ∩ hospital Hospital only 

male 60.5 60.3 62.6% 

female 39.5 39.7 37.4% 

total 100.0 % 

(n=12 668) 

100.0 % 

(n= 21 310) 

100.0 % 

(n= 56 479) 

There is no difference on gender distribution between police-only casualties and 

police hospital casualties, indicating no bias on gender in hospital reporting. 

There is a very slight difference in gender distribution between hospital-only 

casualties and police hospital casualties. This indicates a possible slight bias on 

gender in police reporting. 

Police under-reporting of road casualties and its associated bias factors have 

been studied using a multivariate analysis (Amoros et al., 2006). It was mainly 




shown that : 1) police under-reporting is inversely and strongly associated with 

injury severity, 2) police under-reporting is strongly related to both road user 

type and involvement of a third party. Casualties in crashes involving a third 

party (pedestrian or another vehicle) are more police –reported than those 

without; cyclists are far less police-reported than other road users types. 3) 

police under-reporting is strongly associated with the combination of road type, 

crash environment (“greater Lyon”: inside vs. outside) and police force area. 

Reliability of the linkage 

The reliability of the linkage was assessed on the 2001 data (Amoros et al., 

2007). Once the previously described record-linkage was performed, a survey 

on casualties only identified in the police data was conducted. It consisted in 

going back to the police paper reports and retrieving additional information. In 

particular, names were collected. Since names are available in the road trauma 

registry (since 2000), it was then possible to conduct an additional linkage using 

names on casualties not previously linked. 

First names and surnames (married name if applicable for women) were used. 

Names were compared using the Soundex code to allow for typing, spelling or 

transliteration mistakes. Pairs were pre-selected on matching first name and 

surnames (Soundex coded) ands then linked if also matching on year and 

month of birth, date of crash and crash location. For the 2001 data, the standard 

record-linkage yields 2813 linked casualties, 1322 police-only casualties and 

7823 registry-only casualties. 

The additional record-linkage was able to find 148 additional linked casualties, 

in other words 3.6% of the police casualties, or 1.4% of the registry casualties. 

We checked why these additional linked casualties were not found by our 

“standard” record-linkage: it was either because there were 2 or 3 errors in the 

major linking variables or because the crash location was missing in the 

registry. 

A second assessment of the linkage reliability was provided by estimating the 

number of false positives and false negatives (Amoros et al., 2007). False 

positives are pairs linked whereas corresponding to two distinct casualties; false 

negatives are non-linked pairs whereas corresponding to the same casualty 

(where casualty is defined by an individual and a crash, since an individual can 

be involved in more than one crash). The estimation method was largely 

inspired by two papers : Brenner, 1994 and Brenner and Schmidtmann, 1996. It 

is based on probability computations and approximations of these: probability of 

disagreement in the linking variables (because of errors) for any pair of records 

from the same casualty, and probability of agreement (by chance) for any pair of 

records from distinct casualties. 

On the 2001 data, it was estimated that there were 97 false positives and 396 

false negatives. This corresponds to 2.3% and 9.6% respectively of the police 

records, and 0.9% and 3.7% of the trauma registry records. It further means that 




the number of linked pairs should be increased by 299 (396-97) and the number 

of non-linked pairs decreased by the same amount. 

In conclusion, the number of matches missed by the record-linkage is 

reasonably small. Compared to the total number of casualties of the aggregate 

dataset (police U registry), it is of course smaller. 


As regards to the generalisation to the whole country of correction factors 

estimated at the Rhône county level: the underlying assumption is that the 

police practices (of road casualties reporting and severity classification) are 

homogenous throughout the country. This is supported in France by the 

centralised structure of each police type. As an example it means that the 

degree of police under-reporting of cyclists is assumed to be the same whatever 

the region. By estimating a correction factor for each road user type, the 

variation of the distribution of road user type throughout France can be taken 

into account. The same is true for injury severity; injury severity varies across 

France since it varies between urban and rural areas. Road user type and Injury 

severity are the two under-reporting bias factors used by the common 

methodology. 

There are other important characteristics that display both varying degree of 

police-reporting and varying distributions throughout France : road type, 

urban/rural or police type, and possibly third party involvement. We have shown 

that they are biasing factors in police reporting of casualties. These 

characteristics display different distributions across France: road type 

distribution does vary since some counties do not have motorways for instance. 

Urban/rural distribution or its correlated police type distribution does of course 

vary. Third party involvement distribution may vary since it is correlated with 

traffic density and hence with urban/rural distribution. 

The fact that in this study we take account of only two under-reporting factors 

(those requested for all countries in the project: injury severity) does therefore 

reduce the quality of the estimations. An estimation of the totality of road 

casualties with under-reporting correction factors estimated according to 5 

major bias factors is being conducted; it also includes the estimation of the 

number of non-observed (non-reported) casualties, through the capturerecapture 

approach. It will be published separately. 





Amoros E, Martin J L, Laumon B, 2006. Under-reporting of road crash 

casualties in France. Accident Analysis and Prevention, 38(4), 627-535 

Amoros E, Martin J L, Laumon B, 2007. Estimating non-fatal road casualties in 

a large French county, using the capture-recapture method. Accident Analysis 

and Prevention, May;39(3):483-490 

Brenner, H., 1994. Application of capture-recapture methods for disease 

monitoring: potential effects of imperfect record linkage. Methods of Information 

in Medicine 33 (5), 502-506. 

Brenner, H. and Schmidtmann, I., 1996. Determinants of homonym and 

synonym rates of record linkage in disease registration. Methods of Information 

in Medicine 35 (1), 19-24. 

Clark, D. E., 2004. Practical introduction to record linkage for injury research. 

Injury Prevention 10 (3), 186-191. 




7.4 Study carried out in Greece 

Report prepared by George Yannis, Petros Evgenikos, Antonis Chaziris 

(NTUA) Eleni Petridou, Nikos Dessypris (CEREPRI) 


This report describes the Greek national study on the identification of the road 

accident underreporting level within the framework of Task 1.5 of SafetyNet 

WP1, aiming to estimate the real numbers of road accident casualties, by 

addressing the issue of Police under-reporting. The results of this report will 

allow for the development of appropriate correction coefficients to be applied on 

the Greek road accident data for the estimation of the real number of road 

accident casualties. 

Moreover the development of under-reporting coefficients will allow for 

comparisons between the statistics provided by several EU countries on nonfatal 

injuries. Currently, the only comparable measurement units available in 

CARE, the European Union road accident database with disaggregate data, are 

the numbers of fatal accidents and of people killed, where the degree of 

underreporting is acceptably small in most EU Member States. The same is not 

true for non-fatal accidents and for people suffering non-fatal injuries, therefore 

the numbers of non-fatal accidents and of people seriously and slightly injured 

cannot be compared between the several EU countries. 

Given that the present study aimed to identify possible links between accident 

data derived from the Accident and Emergency Departments of the hospitals 

and the Police road accident data files, it was anticipated that data 

incompatibility problems would be encountered as the data collection systems 

used by the hospital Emergency Departments and the Police differ not only on 

their variable sets but also on the common variables definitions. In order to 

develop the data matrices (as described in the common methodology) to allow 

for the development of the under-reporting coefficients, a methodological tool 

comprising the following steps was adopted. 

The primary target in order to achieve compatibility between the hospital and 

the Police data was to define an appropriate study area to ensure that no 

accident casualties reported by the Police were transferred to any hospital other 

than the one under investigation. This target was achieved by selecting the 

General Regional hospital of Corfu as the source of medical road accident 

casualty data, thus the criterion of a precisely defined "catchment" area was 

met, as all casualties within the island of Corfu are primarily transferred to the 

emergency department of this hospital, even though some more severe injuries 

can be subsequently transferred to better equipped hospitals, mainly in the city 

of Athens. 

In order to perform record linkage between the hospital and the Police data files 

one should primarily identify a set of common appropriate variables to be used 




for the data linking, and ensure that the same definitions concerning these 

variables were adopted in both databases (e.g. fatal injuries are considered 

those in which death has occurred within 30 days following an accident). This 

objective was fulfilled by adopting transformations and/or aggregations in 

specific variables of the hospital database in order to be comparable with the 

respective Police database variables. 

The final step concerned the data matching procedure itself. By using the 

selected variables and values the file with common records in both databases 

(matched cases) was extracted. These records could be grouped by any of the 

selected variables and values allowing their further processing, whereas nonmatching 

cases were grouped in a separate file, which would also be used in 

order to produce the data matrices needed for the elaboration of the correction 

coefficients. 


The study was based on data from two different sources, the Greek Emergency 

Department Injury Surveillance System (EDISS) and the Road Traffic Police 

database. 

Hospital data files 

The Greek Emergency Department Injury Surveillance System (EDISS) was the 

first surveillance system covering all type of injuries and is operated in the 

Emergency Departments of four hospitals in Greece, two of which are located in 

the Greater Athens area (Asclipeion Trauma Hospital and A. Kyriakou 

Children's Hospital), the third being in the county of Magnesia, in Greek 

mainland and the fourth on the island of Corfu. The first hospital is dedicated 

mostly to adult trauma patients, while the second hospital is one of the two 

major Children's hospitals in the Greater Athens area and covers on alternate 

days three quarters of the childhood (


hospital collecting the necessary medical data and meeting all the requirements 

about a predefined "catchment" area is the Regional hospital of the island of 

Corfu, as all road accident casualties on the island requiring medical treatment 

are primarily carried to this hospital (even cases which are subsequently 

transferred to other hospitals). The common methodology for the identification 

of the underreporting level developed within the framework of SafetyNet Task 

1.5 was applied to medical data from this hospital as well as police accident 

data for the prefecture of Corfu, for the period between 1996 and 2003. 

Data collection in EDISS was carried out by specially trained medical staff that 

interviewed patients suffering from any type of injury (or their guardians). A precoded 

questionnaire was used, complemented with a short free text describing 

the injury event. The questionnaire is a modified version of the basic form used 

by all European Union member states, which participate in the European Home 

and Leisure Accident Surveillance System. The Greek version of this 

questionnaire also includes additional variables on traffic and occupational 

injuries, thus the whole spectrum of injuries, by nature, external cause and 

intent is covered by EDISS. The questionnaire covers socio-demographic 

variables, injury characteristics and treatment of the injured individual. For those 

who are eventually hospitalised additional data are also collected. The recorded 

information is entered in a computerised database with continuous data quality 

control system. 

More specifically, the variables recorded in the Greek EDISS are: 

Accident mechanism 

Activity at the time of the accident, Sports 

Location of the accident 

Products involved 

Accident description 

Patient: Age, sex, nationality 

Treatment: Follow-up treatment 

Date and time of attendance 


Injury severity score (ISS) 

Diagnosis: Type of injury (2 possible injuries) 

Injured body part (2 possible parts) 

Administrative information: Country code, Hospital identification number, Patient 

identification number 

First aid 

Place of residence 

Vehicle type (for road accidents) 

Occupation 

Data derived from the EDISS database concern the years 1996-2003, as after 

2003 the data collection was suspended. 




Police data files 

Data collection on road accidents involving casualties (injury or death) in 

Greece is carried out by the Police at national scale since 1963. Whenever an 

injury or death occurs as a result of a road accident, a special force of the Police 

(Department of Road Accidents) carries out an investigation (not only on-site). 

Initially the Police fills-in on the spot an autopsy report. 

Data of the autopsy report are not computerised, but they are processed and 

analysed only at a general level (total numbers, etc). Furthermore, the main 

characteristics of an accident (cause, type, time, day, place etc.) are registered 

in the incident book of the Police department in charge of the accident. 

Data collected by the Police in the appropriate collection form, which is further 

processed by the National Statistical Service of Greece (NSSG), are 

computerised and available for further analyses. The NSSG competent service 

receives by the departmental police offices all road accident collection forms 

filled-in by the police officers in charge of the accident. Two copies are filled-in, 

one of which is submitted to the NSSG, while the other is kept in the local Police 

department. 

The road accident variables recorded by the Police are the following: 

Accident related 

Date (year, month, day of the 

month, day of the week, hour) 

Location 

Area type 

Road type 

Number of casualties 

Number of vehicles involved 

Pavement type 

Weather, lighting and 

pavement conditions 

Accident type 

Vehicle manoeuvre 

Vehicle related 

Vehicle type 

Vehicle nationality 

Vehicle make and model 

First registration year 

Vehicle age 

Number of drivers and 

passengers 

Alcotest data (hour, 

place, results) 

Driving license data 

(nationality, category, 

age) 

Person related 

Road user type 

Gender 

Age 

Nationality 

Use of safety 

equipment 

Casualty severity 

Position in vehicle 

Trip purpose 

All data derived from the Police database concerned the prefecture of Corfu and 

the years 1996-2003, referring to the same area and time period as the EDISS 

road accident casualty records. 

Data processing 

In order to link road accident casualty data, a common subset of variables was 

used to identify the cases to be linked in the two databases. These variables 

were: road user type (driver, passenger, pedestrian, bicyclist); time of 

occurrence (year, month, day); age of the road user (in single years); gender of 




the road user; nationality of the road user and mode of transport. The time of 

the accident was included in the dataset but was not used for the data linking. 

This variable was not considered reliable in the EDISS database, as it was 

recorded while interviewing the casualty/patient (or his/her guardian) at the 

emergency department of the hospital, in some cases many hours after the 

accident occurred. Consequently, time of the accident was used as 

supplementary information for linking cases for which other information was 

missing or the information provided by the other the variables was insufficient. 

The definition of the "mode of transport" variable differs between the EDISS and 

police collection systems, therefore necessary transformations were 

implemented. More specifically, the value "pedestrian" is included in the EDISS 

"mode of transport variable", while it is not included in respective variable of the 

police data file (for pedestrian casualties, the motor vehicle which collided with 

the pedestrian is recorded, and the value "pedestrian" is recorded as value only 

in the "person class" variable). In order to obtain a compatible value set, the 

value "pedestrian" in EDISS was replaced by the vehicle which collided with the 

pedestrian (which was obtained by a third variable included in EDISS). 

Therefore the values included in the EDISS variable "mode of transport" are 

passenger car, motorcycle, moped, bicycle, small truck, large truck and bus. 

As the data files used for the purposes of this study included road accident 

casualties on the island of Corfu, any particularities that arise from the 

characteristics of the selected study area (demographical, traffic related etc) 

were taken into account. More specifically, during vacation periods and 

especially during the summer months, a large number of foreign tourists are 

visiting the island of Corfu, explaining the relatively high proportion of road 

accidents involving foreign people recorded during these periods. This fact was 

exploited to increase the efficiency of the data linking procedure by including the 

variable concerning the nationality of the road user in the set of variables used 

for the data linkage. 

Apart from the common variables and values, some additional variables 

included in the hospital dataset were also included in the data file with the linked 

records. These variables concerned the Length of Stay (LoS) in the hospital, 

and the Abbreviated Injury Scale (AIS score) for each casualty recorded. 

Concerning the AIS scores, these were extracted from the ICD9 scores for each 

casualty recorded in the hospital database. These additional variables were 

used to produce the matrices described in the common methodology, allowing 

for the further processing and the calculation of the under-reporting coefficients. 

The "Length of stay" variable is widely used among hospital databases in 

Europe, although it is not defined in the same way throughout the countries. The 

Greek definition of length of stay is based on whether the date has changed 

while the patient was being hospitalised. For example, if the patient was 

hospitalised while the date has not changed (e.g. from 05:00 to 19:00 at the 

same day), or if the patient was hospitalised while the date changed only once 

(e.g. from 18:00 to 09:00 in the next morning), then the length of stay would be 




1. If the patient stayed hospitalised while the date changed twice, the length of 

stay is 2, etc. 

Therefore, the Greek definition of the LoS is not necessarily related to the nights 

spent, but with the change of date while the patient was being hospitalised. It is 

also evident that the length of stay will be 1, even for some cases where the 

patient did not spend the night in the hospital. 

Any ethical issues that arise from the use of person identification information 

related to road accident casualties were also taken into account. Variables 

concerning personal information (casualty names, personal addresses, date of 

birth etc) were not included in the data files used for the record linkage. Any 

information that could be used for the identification of a casualty at personal 

level is considered confidential and could not be used for research purposes. 

The data linking was carried out in two distinct phases, each one concerning 

fatalities and injuries respectively. Concerning non-fatal injuries, the severity of 

the accident is defined differently in the hospital and the police databases. More 

specifically, the police defines injury severity using personal judgement of the 

police officer filling-in the accident form, while the hospitals define injury severity 

on the basis of the days of hospitalisation. Thus, many injuries defined as 

serious in the hospital database could be identified as slight by the police and 

vice versa. Consequently, linking the police serious injuries with the hospital 

serious injuries and the police slight injuries with the hospital slight injuries could 

result an important loss of potentially linked cases which were recorded 

according to a different injury severity definition in the two databases. On that 

purpose, the data linking was carried out using the total number of injuries 

(serious and slight together) in the two databases. 

After completing the data linkage for all injuries, an additional task was 

performed in order to identify the proportion of casualties for which the accident 

severity was differently reported (serious in the one database and slight in the 

other) in the hospital and the police data files. This proportion was relatively 

high (only 47,8% of the linked records' injury severity agreed between the two 

databases) verifying the choice to link aggregated injury type data files. 


The Regional Hospital of Corfu was the only one out of the four hospitals 

collaborating in EDISS meeting the clearly defined "catchment" area 

requirements. Indeed, all road traffic casualties occurring on the island and 

requiring medical treatment are initially assessed in the Emergency Hospital 

Department, even the most severe cases which are subsequently transferred in 

other hospitals for specialised care. The linking procedure was applied on data 

concerning the Emergency Medical Department of the Hospital and the Road 

Traffic Police data for the prefecture of Corfu between 1996 and 2003. 

The present study aimed to link hospital and police road accident casualty data 

files. The two data files included a different series of variables and values, 




therefore all transformations / aggregations required to obtain the data grouped 

by the same variables (as already described in "Data description") were 

adopted. After all necessary adjustments have been made, all data grouped by 

the linking variables were included in a single file which was the input for the 

linking process. 

In order to identify the common records between the two datasets specialised 

statistics software was used. The input file included both hospital and police 

records for the whole series of years grouped by the variables used for the 

linking. The utility "identify duplicate cases" was used to identify the matched 

pairs of records in the data file, based on the given set of linking variables. 

The result of this procedure grouped the data file in the following way: 

- A case from one file matched with one case from the other file (perfect 

matches). 

- A case from one file matched with more than one cases from the other file. In 

that case the additional variables (such as the time of the accident) were used 

to identify the most likely among the various matches. The selection was done 

manually. 

- Two (or more) cases from the same file were matched. These cases 

concerned accidents with identical characteristics in the same database, 

therefore they were not taken into account in the data linkage. 

- More than one cases from one file were matched with more than one cases 

from the other file. Once again, additional variables were used to identify the 

most likely pairs among the various possible matches. 

- Non-matched cases. 

During the linkage process, however, it was revealed that some values of the 

common set of variables used for the cross-checking did not match for several 

records in both datasets, although they seemed to refer to the same casualty. 

Inconsistencies between the values could be attributed to incorrect reporting 

from the part of the patient in the case of the hospital data or to misjudgement 

on the part of the person completing the collection form, in the case of the police 

data or simply due to an error while processing the data. In order to include 

these cases to the linked data file one of the following two approaches were 

adopted: 

1) Repeating the linking process using each time a different subset of less 

common variables as the criterion for linkage in order to examine a broader 

range of possible matches. For example, when the nationality of the road 

user was unknown in the police database and known in the EDISS database, 

excluding the nationality variable from the matching subset of variables 




would allow inclusion of the case in the possible matching groups, based on 

the matching results of the remaining variables 

2) Adopting less strict rules while linking the data, thus, allowing a tolerance 

interval for the values of certain variables. Consequently, some cases could 

be matched although slight differences on the values of one linking variable, 

was observed, while the rest of the variables provided sufficient evidence for 

linking the cases. The variables for which such tolerance intervals were 

allowed were the following: 

Date of the accident (a tolerance interval of one or two days before and/or 

after the actual date of the accident could be allowed). 

Nationality of the road user - only in cases where the nationality was 

recorded as "Albanian" in the police data file and as "from North Epirus" in 

the EDISS data file. North Epirus is a region officially within the borders of 

Albania with a minority of Greek population, which is recorded using a 

separate value in the EDISS database. A large number of people from North 

Epirus are living on the island of Corfu. 

Age of the road user (a tolerance interval of 5 years was adopted in some 

cases, in order to include possible rounding of the age in the databases). 

Mode of transport variable (some cases where "moped" was recorded in 

the one data file and "motorcycle" in the other were linked, when sufficient 

evidence from the rest of the variables was provided. Such links were 

established also for a few cases where small truck was recorded in the one 

data file and passenger car in the other). 

Each time the record linkage was taking place, a set of linked hospital-police 

records was being copied into a separate file containing the linked cases. All 

other records were entering the next iteration. 

After performing several times the same procedure, using each time different 

variable sets all possible matches were identified. The iterations were stopped 

when the set of variables used for the data linking was judged to be small 

enough therefore unable to provide sufficient data for a pair of records to be 

linked. 

Obviously, when less strict rules for the record linking were used, the manual 

checking for each pair of records was more important in order to avoid linking 

irrelevant records. The records left in the end were the non-matched cases. 

These data were transferred into separate files for the hospital and the police 

data. This file contained road accident casualty records, recorded by the police 

and not recorded by the hospital and road accident casualties recorded by the 

hospital and not recorded by the police. 




7.4.4 Results 

Summary results of the overall linking results are provided in this section. 

Moreover, some additional results concerning the presentation of the data 

linkage in disaggregate form are presented in the following Tables. More details 

concerning the main results of the Greek national study, the data matrices used 

for the extraction of the under-reporting coefficients, as well as data concerning 

the length of stay or the AIS scores for the casualty records, are provided in the 

main report. 

The main results of the record linkage between the hospital road accident 

casualty database records of the General Regional hospital of Corfu and the 

police records referring to the respective "catchment" area are summarised in 

the following Table 63. 

Table 63: Summary of linking results 

Fatalities Injuries 

Records in Hospital database 97 11.267 

Records in Police database 172 1.910 

Linked cases 91 1.262 

Not linked hospital records 6 10.005 

Not linked police records 81 648 

Total records (real number) 178 11.915 

During the eight year study period (1996-2003) 11.364 road accident casualties 

contacted the Corfu Hospital Emergency Medical Department and out of them 

97 (0,9%) died on arrival to the hospital or during hospitalization. For the same 

period 2.082 road accident casualties were reported by the police and out of 

them 172 (8,3%) were declared as deaths. 

Only 6 out of the total 178 cases (3,3%) concerning fatalities were not reported 

in the police database, while the respective percentage for injuries is 

significantly higher (10.005 out of 11.915, corresponding to a proportion of 

84%). Additionally, a high percentage of road accident fatalities are not reported 

in the hospital database, most likely due to a large number of deaths at the 

accident site, never contacting the emergency department of the hospital. The 

number of non fatal injuries reported by the Police and not reported in the 

EDISS database is significantly lower (648 out of 11.915, 5,4%). 

The results of the linking procedure are presented below in schematic form. 




Results of data linkage for fatal 

injuries 

Results of data linkage 

for non-fatal injuries 

The results of the linking procedure were further processed in order to provide 

the figures concerning the hospital and the police casualty reporting in 

disaggregate form. More specifically, the following Tables present the linking 

results by the casualty's age and gender, mode of transport (for pedestrian 

casualties the linked vehicle was recorded) and nationality. 

Table 64: Fatalities by age groups and gender 

Linked EDISS only Police only 

Age groups Male Female Total Male Female Total Male Female Total 


Table 65: Injuries by age groups and gender 

Age 

Linked EDISS only Police only 

groups Male Female Unknown Total Male Female Unknown Total Male Female Unknown Total 


respective number for any other mode of transport, both fatal and non-fatal 

accidents. More specifically, almost 63% of the EDISS only non-fatal injuries 

refer to motorcycle related accidents, while the respective figure for passenger 

cars is approximately 15%. 

Table 67: Fatalities and injuries by nationality 

Fatalities 

Injuries 

Nationality Linked 

EDISS 

only 

Police 

only Linked 

EDISS 

only 

Police 

only 

Greek 67 6 57 958 7.532 342 

Albanian 5 0 5 72 535 46 

Italian 3 0 3 26 321 30 

British 6 0 8 96 392 97 

German 4 0 5 40 388 39 

Other 6 0 3 70 837 94 

Total 91 6 81 1.262 10.005 648 

Table 67 presents the linking results concerning both fatalities and non-fatal 

injuries by nationality. These results could prove useful for identifying 

differences in the underreporting levels between natives and foreigners. 

Excluding Greek casualties, the highest proportion both in fatal and non-fatal 

injuries is observed for the British people, followed by the Albanians. 

In summary, the results presented in these Tables provide a more detailed view 

of the road accident casualty data linkage between the EDISS and the police 

databases for the prefecture of Corfu. It is evident though that disaggregate 

data concerning fatalities refer to a small number of records and their further 

division in subgroups (age, gender etc) cannot be used for further calculations 

due to the small size of the sample. Nevertheless, the non-fatal injury figures 

present a great potential for further analyses as the sample size is considerably 

larger. 


The methodology used to link road accident casualty data between the hospital 

(EDISS) database and the respective police data files is based on using a 

statistics utility for identifying duplicate cases within a single data file. The 

procedure used for the data linking as described in Section 7.4.3 is not fully 

computerised and its correct application strongly depends on the manual 

checking of the linked records. Although it provides reliable results (every linked 

record pair is manually checked and it is very unlikely that a matched pair will 

not be found) it would be very difficult to implement such a procedure on large 

data files (containing a long time series and a bigger area of study), as the 

manual checks of the record matches would be extremely time consuming. In 

the present study the sample consisted of 12.000 records approximately, most 

of which could not be linked to another record (only 1.262 matched cases). 




The input file used within this study consisted of both hospital and police 

records for the whole series of years (1996 - 2003) grouped by the six variables 

used for the linking (road user type, date of occurrence, age of the road user, 

gender and nationality of the road user and mode of transport) and other 

variables which would serve either as supportive information on whether a 

record pair should be linked or not (for cases where more than two records were 

linked using only the six variables) or as information to be included in the data 

matrices required for the extraction of the correction coefficients (as described 

in the common methodology). The inclusion of several variables in the common 

data file made possible to extract the linked and not linked data files in a 

disaggregate form, using each time a different variable (or set of variables) to 

describe the data, therefore it can allow for the development of more detailed 

correction coefficients. 

A study concerning road accident casualty under-reporting coefficients at a 

national level should always be based on appropriate data, meaning that the 

sample should not only have a statistically significant size (e.g. for a series of 

years) but also being representative for the selected country. Therefore, in order 

to evaluate the results and the conclusions from this study, any restrictions that 

may arise from the collected data should be taken into account. The data used 

for the present study concerned the island of Corfu, representing only a 

proportion of the overall road accident casualties in Greece, being also an ideal 

"catchment" area for matching hospital and police data. The high presence of 

tourists during the summer period does not necessarily affects the matching of 

the two files, however further investigation could be proved useful. 

This study revealed not only the great potential for linking hospital and police 

data but also the need for further investigation of such links in order to come up 

with larger samples and thus more complete results. 





Dessypris N., Petridou E., Skalkidis Y., Moustaki M., Koutselinis A. and 

Trichopoulos D. (2002). Countrywide estimation of the burden of injuries in 

Greece: a limited resources approach. J Canc Epidem & Prev 2002; 7: 123-129. 

Petridou E, Gatsoulis N, Dessypris N, Skalkidis Y, Voros D, Papadimitriou Y 

and Trichopoulos D (2000). Imbalance of demand and supply for regionalized 

injury services: a case study in Greece. International Journal for Quality in 

Health Care; Vol 12; Issue 2; p.115-113. 




7.5 Study carried out in Hungary 

Report prepared by Dr. Gábor Merényi, Olivér Zsigmond, Árpád Tóth, Dr. 

Péter Holló (KTI) 

7.5.1. Introduction 

In Hungary, all hospitals report on their discharged patients to Gyógyinfok, the 

firm engaged in data management at the National Health Insurance Fund. The 

report includes the discharged patients’ date of birth, the time of their 

hospitalisation and discharge, their gender as well as the social insurance 

identification number (TAJ). On the basis of this identification, the case may be 

followed even if the patient is transported to another hospital. In Hungary, too, 

disease (injury) recording is implemented in accordance with the BNO-10 (ICD- 

10) code system. In case of accident, the accident cause has also to be 

registered in the BNO-10 code system. Traffic accidents, and only this type of 

accidents, all begin with „V”, therefore they can be easily selected from the 

database. 

From the aspect of the task, unfortunately the Gyógyinfok Database has several 

disadvantages: 

- Because of data protection aspects, the TAJ is not included into the 

database of the police, therefore the person cannot be identified as simply 

as that 

- Because of different financing, the attended outpatients are recorded in 

another database, and until 2006 it was not compulsory to code the cause 

of the accident, therefore, traffic accidents cannot be retrieved from this 

database. 

- The BNO (ICD) –10 codes are not suitable for specifying the AIS (MAIS) 

severity scale demanded by the task with the software available with the 

database. 

Because of the above reasons and for the sake of more precise data collection, 

the whole national database has not been used, but all data from one selected 

trauma centre were analysed and compared with the police data. 

7.5.2. Description of data sources 

This centre is the Károlyi Sándor Hospital in Budapest. This hospital is one of 

the 4 Regional Trauma Centres in Budapest. All the injured adults of the traffic 

accidents occurring in five northern districts of the Capital (III.,IV.,XIII.,XIV.,XV.), 

and in the approximately 25-30 settlements of the agglomeration, as well as on 

the northern main roads (2,2A,2B,3,M3,10,11) leading to the Capital are brought 

here on five days of the week (Tuesday-Wednesday-Friday-Saturday-Sunday) 

throughout 24 hours. 

From the hospital’s computerised database, the data of all the patients were 

processed who had been injured in traffic accidents and hospitalised in the 




hospital’s department in the period between 1 August 2004 and 31 January 

2006. 

The following data have been selected in order to be compared with the police 

database and to be processed: 

The patient’s date of birth, gender. 

Accident time, location. Type of the vehicle involved in the accident, the 

role of the injured, the accident’s mechanism (if there was any reference 

in the data on health). 

The attendance case of the injured has been classified on the basis of 

hospitalisation: out-patient, overnight, 1-3 days, or over 3 days attendance. 

According to AIS (Abbreviated Injury Scale) injury severity has been 

determined for 6 body regions (head & neck, face, chest, abdomen-spine, 

extremity, external). Accordingly severity of different body regions’ injuries are 

classified from 0 to 6. In compliance with the categorisation of the American 

Association for the Advancement of Automotive Medicine, the different scales 

are follows: 

1 – Minor 

2 – Moderate 

3 – Serious 

4 – Severe 

5 – Critical 

6 – Unsurvivable 

MAIS (maximum of the AIS), the highest AIS value has been determined for 

each injured person. The AIS values were coded directly by medical staff. 

ISS (Injury Severity Score) has been calculated in each injury case separately. 

Calculation of this generally applied indicator: the total amount of the AIS 

squared value calculated for the three most seriously injured body regions . This 

value may be 0-75. (If on some body region the AIS=6, then the ISS will be 

automatically 75.) 

Furthermore, the number of persons deceased within 30 days in hospital have 

been recorded. 

7.5.3. Description of the linking process 

Starting data: 

• ACCIDENT data set: 

KSH (police), 2005CA.dbf 

28783 records, out which 2665 records within catchment 




• HOSPITAL data set 

KORHAZ6.dbf 

1294 records 

(only in the catchment, adult injured, not complete (e.g. falling by cycle not 

included, only 5 days of the week are concerned) 

The data sets, clarified and coded were available in this item only for 2005. In 

the hospital the data have been coded for a longer period, as described in 7.5.2, 

but the problem is that the official data for 2006 are available only as of mid 

2007. Therefore, at the time of the linkage we could use mainly the data for 

2005. 

Matching 

According to its content, the HOSPITAL data set has been divided into two 

parts: 

- K1-complete: the key fields (DATE, AGE, GENDER, PLACE) 

determinant from the aspect of matching are available 

- K2-incomplete: some parts of the data are missing 

The ACCIDENT data set has been divided into two parts: 

- B1-in the catchment: if the accident occurred in the catchment area of the 

hospital 

- B2-outside the area 

The process of matching was implemented in three successive steps: 

1) K1 – B1 (result: 474 data-series agree out of 1012 cases) 

2) K2 – B1 remainder (result: 22 data-series agree out of 282 cases) 

3) (K1+K2) remainder – B2 (result: 4 data-series agree out of 798 cases) 

matched total: 500 events out of 1294 cases 

Breakdowns: 

Both data sets include 

Only the HOSPITAL data set includes 

Only the ACCIDENT data set includes (in the area of catchment) 

500 cases 

794 cases 

2165 cases 

The process of matching was carried out by computer, not manually. 

Experience showed that the difference allowed in the accident time (±1 hour) 

was too limited: the values in the hospital database are rather uncertain as they 

are based on declaration, or assessment. As a follow-up to this work, this 

tolerance should be increased, then the degree of matching of the two 

databases would certainly be greater. 




Conclusions concerning the linking process 

In the first step of matching, the agreement of gender + age + date + time + site 

has been investigated. The criterion of congruency is the precise agreement of 

the gender and the date, in the case of the age ±1 year, in the case of the time 

Ki-Bi


7.5.4. Results: 

During the 18 months of the survey 2107 persons injured in traffic accidents 

attended the hospital. 

Role: driver 1307 

Front seat passenger 52 

Rear seat passenger 75 

Unknown passenger 342 

Pedestrian 331 

TOTAL 2107 

Nursing days: Outpatient 1101 52,3% 

1 day 54 2,6% 

1-3 days 465 22,1% 

>3 day 487 23,1% 

TOTAL 2107 

MAIS 1 985 46,7% 

2 688 32,7% 

3 319 15,1% 

4 67 3,2% 

5 40 1,9% 

6 8 0,4% 

TOTAL 2107 

Deceased 9 0,42% 

ISS 0-15 1942 92,2% 

16-30 117 5,6% 

31-45 34 1,6% 

46-60 6 0,3% 

61-75 8 0,4% 

TOTAL 2107 

Motorcycle 373 

Passenger car 946 

Bus 32 

Heavy vehicle 17 

Trailer 1 

Special vehicle 0 

Tramway 2 

Trolleybus 5 

Suburban railways 0 

Train 0 

Bicycle 356 

Moped 40 

Animal 0 

Other 4 

Pedestrian 331 

TOTAL 2107 




Relation between vehicle type, injury severity and the period of stay in the 

hospital 

Nursing days 

vehicle MAIS outpatient overnight 1-3 days >3 days Summary 

car occupant 1 417 4 69 9 499 

2 84 14 149 50 297 

3 1 1 22 79 103 

4 1 29 30 

5 14 14 

6 3 3 

car occupant Sum 502 22 241 181 946 

motorcyclist 1 126 2 21 2 151 

2 63 1 36 24 124 

3 6 1 13 57 77 

4 11 11 

5 9 9 

6 1 1 

motorcyclist Sum 195 4 70 104 373 

pedal cyclist 1 156 4 8 4 172 

2 79 2 28 12 121 

3 9 6 6 34 55 

4 4 4 

5 1 3 4 

pedal cyclist Sum 244 12 43 57 356 

pedestrian 1 72 3 28 9 112 

2 30 7 54 23 114 

3 1 6 61 68 

4 1 20 21 

5 12 12 

6 3 1 4 

pedestrian Sum 103 14 88 126 331 

other 1 41 9 1 51 

2 14 2 12 4 32 

3 2 2 12 16 

4 1 1 

5 1 1 

other Sum 57 2 23 19 101 

1 part 812 13 135 25 985 

2 parts 270 26 279 113 688 

3 parts 19 8 49 243 319 

4 parts 1 1 65 67 

5 parts 1 39 40 

6 parts 6 2 8 

Summary 1101 54 465 487 2107 




Relation between vehicle type and the severity scale of the body’s 

regional severity: 

Head&Neck car occupant motorcyclist pedal cyclist pedestrian Other Summary 

1 395 30 53 80 24 582 

2 199 41 41 78 13 372 

3 6 5 8 6 1 26 

4 5 3 1 7 1 17 

5 11 5 4 10 1 31 

6 1 3 4 

Summary 616 85 107 184 40 1032 

Face car occupant motorcyclist pedal cyclist pedestrian other Summary 

1 61 10 9 14 1 95 

2 14 6 7 20 1 48 

3 2 2 1 2 7 

4 1 1 

Summary 77 18 17 37 2 151 

Chest car occupant motorcyclist pedal cyclist pedestrian other Summary 

1 145 26 19 33 12 235 

2 101 29 11 25 5 171 

3 33 10 6 16 2 67 

4 7 2 2 11 

5 1 1 

6 2 1 3 

Summary 288 68 36 77 19 488 

Abd.Spine car occupant motorcyclist pedal cyclist pedestrian other Summary 

1 78 29 14 29 6 156 

2 13 5 1 3 4 26 

3 17 16 2 7 2 44 

4 9 5 1 4 19 

5 2 1 3 6 

6 1 1 

Summary 120 56 18 46 12 252 

Extremity car occupant motorcyclist pedal cyclist pedestrian other Summary 

1 151 112 101 76 30 470 

2 50 55 64 40 12 221 

3 65 64 44 67 12 252 

4 18 6 2 19 45 

5 4 3 3 10 

Summary 288 240 211 205 54 998 

External car occupant motorcyclist pedal cyclist pedestrian other Summary 

1 364 166 198 155 46 929 

2 78 61 28 61 5 233 

3 5 2 2 9 

Summary 447 229 226 218 51 1171 




Relation between vehicle type and injury severity 

MAIS 

Vehicle 1 2 3 4 5 6 Summary 

car occupant 499 297 103 30 14 3 946 

motorcyclist 151 124 77 11 9 1 373 

pedal cyclist 172 121 55 4 4 356 

pedestrian 112 114 68 21 12 4 331 

other 51 32 16 1 1 101 

Summary 985 688 319 67 40 8 2107 

7.5.5. Conclusions 

More than half of the injured (52,3%) have been attended as outpatients. 23,1% 

of the injured had been hospitalised for more than 3 days. 

Accordingly, 80% of the injured suffered minor or moderate injuries (contusions, 

lesions, simple fractures, concussion). 

5,5% of the injured had very severe, critical or fatal injury. 9 injured (0,4%) 

deceased. Of course, the latter means a death in the hospital and within 30 

days. 

Cycle accidents are frequently not included in the police database, and also the 

hospital data are deficient (e.g. the accident site), since very often neither the 

injured nor the doctor considers the falling with a bicycle on the public road as a 

traffic accident. In these cases the injured does not call out the police, and 

usually the hospital does not inform the authorities either. At the same time, 

relatively serious injuries may also occur in this category (e.g. shoulder 

fractures, dislocations). 

Head, chest and abdominal injuries were in higher proportion in the case of 

injured car occupants, whereas pedestrians, motorcyclists and cyclists suffered 

more frequently of extremity-injuries. 

The number of persons injured in passenger cars was almost three times higher 

than that of motorcyclists, cyclists and pedestrians separately. The accident 

victim rate in the case of these three latter groups was the same. Against them, 

the number of injured in other vehicles was insignificant. 

Nevertheless, the number of critical and fatal cases involving pedestrians was 

the same as that concerning the car occupants (i.e. in proportion threefold 

majority). The rate of very serious and fatal injuries were also higher in the case 

of motorcyclists than motorists. 

Accordingly, while in the case of car occupants the number of injured attended 

as out- and overnight patients were in majority in comparison with those 

hospitalised for several days, this rate is reversed if pedestrians are concerned, 




because almost twice as many injured had been attended in hospital for several 

days. 

Representativeness of the results 

We have compared the different injured categories for the catchment’s area of 

the hospital and for the whole capital (Budapest). The distributions of the injured 

persons are as follows: 

Vehicle 

Sample 

Budapest 

categories injured % injured % 

Motorcycle 260 20,09 428 8,36 

Car 546 42,19 2619 51,16 

Bicycle 236 18,24 232 4,53 

Pedestrian 194 14,99 1315 25,69 

Other 58 4,48 525 10,26 

All 1294 99,99 5119 100,00 

It can be seen, that the sample was not representative even for Budapest. The 

percentages of injured motorcycle and bicycle riders are greater in the sample 

than in Budapest, and the percentages of injured car occupants and pedestrians 

are less. The representativeness is even less regarding the whole country. 

The following are the differences between a trauma centre in the capital and 

that one in a county hospital: 

a) No child under 14 will be admitted to the former (admission is usually 

possible to country centres) 

b) There is no non-stop inspection over the 7 days of the week (5 days were 

investigated by the hospital) 

c) There are out-patients’ district dispensaries where the patient can 

present himself (in the country they are mostly substituted by surgeries 

linked to hospitals, so that they are also registered in the hospitals’ 

database) 

Studies in the future 

Due primarily to likely lack of the representativeness of the sample of matched 

data, it is proposed to: 

a) involve one countryside centre into the studies 

b) carry out a prospective study in order to make possible recording of the 

predetermined data. (The character of the study detailed here – as it 

turns out from the particulars – has been retrospective.) 

Another possibility can also be imagined for future studies. Besides taking into 

consideration the aspects of the protection of the personal data, an attempt 

should be made to study the databases of the central public health 

(GYÓGYINFOK) and of the police (KSH) on the basis of BNO-10. The coding of 




the accident cause is already compulsory for the hospital and outpatient cases, 

thus all the codes starting with “V” correspond to traffic accidents. Moreover, 

also the role of the injured can be derived from this. The TAJ number is 

personal data, but a crosscheck could be made on the basis of the date of birth 

and the starting date of hospitalisation. 

Combined Police and Medical data set - HUNGARY 

LENGTH of STAY matrix 

Police coding 

Road user type Length of Stay Fatal Serious Slight not matched Total 

Driver out-patient 11 79 90 

Overnight 1 4 5 

1-3 days 19 54 73 

>3 days 74 21 95 

not matched 39 255 802 1096 

Passenger-front out-patient 1 39 40 

Overnight 1 3 4 

1-3 days 3 23 26 

>3 days 12 5 17 

not matched 11 67 303 381 

Passenger-rear out-patient 2 19 21 

Overnight 1 1 2 4 

1-3 days 14 14 

>3 days 10 1 11 

not matched 4 52 254 310 

Pedestrian out-patient 1 16 17 

Overnight 1 2 3 6 

1-3 days 7 27 34 

>3 days 30 13 43 

not matched 29 111 238 378 

not matched out-patient 460 460 

Overnight 21 21 

1-3 days 175 175 

>3 days 138 138 

Total 86 659 1920 794 3459 




MAIS matrix 

Police coding 

Road user type MAIS Fatal Serious Slight not matched Total 

Driver 1 10 83 93 

2 27 59 86 

3 54 14 68 

4 11 1 12 

5 3 1 4 

not matched 39 255 802 1096 

Passenger-front 1 2 38 40 

2 2 27 29 

3 10 5 15 

4 2 2 

6 1 1 

not matched 11 67 303 381 

Passenger-rear 1 1 21 22 

2 2 14 16 

3 4 1 5 

4 4 4 

5 2 2 

6 1 1 

not matched 4 52 254 310 


2 8 27 35 

3 18 4 22 

4 7 7 

5 3 3 

6 1 1 2 

not matched 29 111 238 378 

not matched 1 377 377 

2 283 283 

3 102 102 

4 18 18 

5 14 14 

Total 86 659 1920 794 3459 




7.6 Study carried out in the Netherlands 

Report prepared by Niels Bos (SWOV) 

Within the framework of the Safetnet project, a study was carried out into linking 

hospital data and police data on traffic casualties in The Netherlands. Both 

databases contain a number of key variables that enables matching of patients 

in one database with casualties in the other. This was done in order to estimate 

the level of underreporting and to verify the severity that the police assigns to a 

casualty. Police estimated severity not always appeared to be of the same 

severity in the hospital database, as was investigated by using the MAIS scores. 

Correction factors have been determined which, applied to the CARE accident 

data, should describe the development of real numbers of severely injured 

casualties. These factors are different by mode of transport. However, underreporting 

and differences in distribution over relevant variables, such as year of 

accident and age of the casualty, prevent outcomes being very reliable. 

At the moment, linkable Dutch hospital data is available at SWOV for the years 

1997-2005 in a consistent database using uniform coding with ICD9-cm. In this 

study the years 1997-2003 are used. Linkable police data is available for the 

years 1976-2005, of which the CARE database contains the years 1991-2003. 


In this study the real number of hospitalised road casualties is derived from 

linking a hospital discharge file to an accident file of police reported road 

accident casualties. Based on the intersection of "distance based matches" and 

both rest files, an estimate is made for the number that was not reported in 

either database. This leads to a total number of casualties that can also be 

examined on their medical severity. This severity (MAIS) was determined on all 

injury codes that are in the hospital file and form an objective scale of severity. 

The severity will also be expressed in Length of Stay. 

The linking study is extensively described in Dutch, in Reurings, Bos & Van 

Kampen (R-2007-8). In this report only an overview is presented and for details 

reference is made to the SWOV study. 

For a correct assessment of developments in road safety and of the effects of 

road safety measures and interventions the availability of reliable figures on 

traffic injury outcomes is crucial. In recent years, the governments of many 

countries have started to set targets for their future national road safety. These 

targets are usually expressed in terms of numbers of traffic injury outcomes. 

Setting up realistic targets also depends to a large extent on the availability of 

accurate figures on current traffic injury outcomes. In the Netherlands, as in 

many other countries, fatal outcomes in traffic accidents are registered quite 

well, but this is generally not the case for hospitalised road traffic casualties. 




In order to determine the degree of underreporting of hospitalised road traffic 

casualties in Dutch police reports, an extensive study was carried out in which 

the records of two well-established databases were compared. The first 

database is the so-called AVV database of police-reported road accidents, 

which contains extensive information on the accident site, and on the vehicles 

and casualties involved in the accident. The second database is the so-called 

Prismant-Hospital Discharge database. This is the official database containing 

information on all patients admitted into Dutch hospitals. In this database 

reporting is mainly concerned with medical issues. 

The police database only contains information on road traffic casualties and 

accidents, and is known to be incomplete (Polak, 1997). Moreover, both 

hospitalised and non-hospitalised casualties are recorded in this database. The 

hospital database, on the other hand, only contains information on hospitalised 

patients and is known to be (almost) complete. However, many of these 

hospitalised patients are not road traffic casualties, while road traffic casualties 

are not always recorded as such. 

The aim of the present study was to asses the actual number of non-fatally 

injured but hospitalised road traffic casualties from these two databases. Thus, 

the group of interest consists of all casualties of a road accident (according to 

the international definition) who have been inpatients in a hospital as a result of 

the accident, and who did not die within 30 days after the accident. The 

determination of the maximum AIS score allows us to set a boundary in order to 

identify real seriously injured road traffic casualties among them. In contrast, 

road traffic casualties who die within 30 days of the accident (whether 

hospitalised or not) belong to the group of fatally injured road traffic casualties. 

Considering the nature of the two Dutch databases, the seriously injured road 

traffic casualties can be classified into four subgroups: the hospitalised road 

traffic casualties recorded in both databases, those recorded in only one of the 

two databases, and those that are missing in both databases. Therefore, the 

group that has to be recovered consists of the four cells enclosed with double 

lines in Table 68, which contains all possible combinations. 

Table 68: Possible combinations of presence or absence of hospitalised 

road traffic casualties in police and hospital databases. 

In police database 

Not in police 

database 

Not a hospitalised 

road traffic 

casualty 

In hospital 

database 

In both 

databases 

Only in hospital 

database 

Hospitalisation 

not caused by a 

road traffic 

accident 

Not in hospital 

database 

Only in police 

database 

In neither one 

database 

Not a hospitalised 

road traffic casualty 

Road traffic casualty 

but not hospitalised 




To identify these four subgroups of the hospitalised traffic road casualties, a 

linking procedure was applied which compares a number of key variables 

contained in the two databases on a record by record basis. 

The linking procedure used in the present report is well suited for situations 

where unique identifiers like a personal ID number are not available for linking. 

A generalised distance function is defined which quantifies the similarity 

between pairs of records in the two databases. This quantified similarity can be 

used to assess the probability of the correctness of a match: the smaller the 

distance, the higher the probability that the two records refer to the same 

individual. 

The linking procedure is probabilistic and conjunct. It is probabilistic because 

discrepancies between records are tolerated, including missing information. The 

procedure is conjunct because it simultaneously compares all the records in the 

first database with all the records in the second database 3 , and therefore only 

requires two passes through the data. Other probabilistic methods for linking 

police and hospital records like GIRLS (Generalised Iterative Record Linkage 

System) are disjunct. In GIRLS records are grouped according to their scores 

on a key variable, and compared within each group. Matched records are then 

removed from the linking process. In the next pass the remaining records are 

grouped according to their scores on another key variable, and again compared 

within each group, etc., etc. This repeated grouping of records and removal of 

linked records creates order effects, and requires many passes through the data 

(see, e.g., SWOV, 2001 or Rosman et al., 1996). 

In the conjunct method two records are matched when they are each other’s 

nearest neighbours in terms of distance or similarity. Moreover, the difference 

between the distance of a matched pair of records and the distances to their two 

next best neighbours is used to quantify the selectivity (or exclusiveness or 

uniqueness) of the match. This selectivity measure provides a second 

diagnostic for the probability of correctness of a match. 

The police and hospital databases, as well as the key variables used for linking, 

first are described in section 7.6.2. Then, in section 7.6.3 a generalised distance 

function is introduced which quantifies the similarity between records in the two 

databases, even if information is missing in one or both records. The linking 

procedure itself is described and applied to the Dutch police and hospital 

databases. In section 7.6.4, results are analysed in terms of the severity 

(Maximum AIS) and Length of Stay in hospital (LoS). In this section also the 

results that are required within the SafetyNet project are given. Finally, in 

section 7.6.5 conclusions will be drawn. 

3 Except for the time frame. Records are only compared in our study if they are within -1 to +3 

days apart. 





7.6.2.1 The police database 

In the police database for road accidents (VOR) all accidents in the Netherlands 

in which at least one person is injured are collected. This statistic should cover 

100% of all relevant cases in The Netherlands, but we know that hospitalised 

casualties are underreported by about 40% (Polak, 2000). The police collects 

the data by filling in a paper report form. All these papers are collected by the 

Ministry of Transport (AVV) and transformed into an electronic format. Persons 

who successfully committed suicide (confirmed) are excluded from the 

database. 

For the linking procedure we used data from the years 1997-2003. In these 

years, 324.717 persons were either injured or killed in a road accident according 

to the police. As in most of the European countries, the 30 day definition for 

fatalities is used: If a person dies within 30 days after the accident the person is 

counted as a road accident fatality. Uninjured persons involved in road 

accidents where excluded from the linking process. The casualties were coded 

by severity, of which only category 6 is defined as hospitalised: 

Table 69: Number of casualties in the police database by severity. 


cases 

code description 

(1997-2003) 

Fatalities 0 Killed on the spot 4.431 

1 Killed later the same day 1.435 

2 Killed one day after 537 

3 Killed 2-5 days after 452 



Hospitalised 6 Hospitalised 79.984 

Slightly 

7 Transferred to hospital, 

90.738 

injured 

not hospitalised 

8 Transferred to hospital, 

11.304 

hospitalisation unknown 

9 Not transferred to hospital 126.402 

10 Transfer and hospitalisation unknown 8.880 

SUM 324.717 

7.6.2.2 The hospital discharge database 

In the hospital discharge database (LMR) administrative and medical data of inpatients 

of all Dutch hospitals are collected. This statistic is supposed to cover 

100% of all relevant cases. Outpatients are not recorded in this database. Data 

is collected by the hospitals and transferred to Prismant which prepares the 

hospital discharge files. 




We received records for patients that were injured in a traffic accident, but also 

of cases in which the cause was unknown (E-codes E800-E829+E928+ 

E958+E988). In this database the main diagnosis from hospital treatment is 

recorded. This is determined afterwards, at the moment that the patient is 

discharged. The hospitals use the international classification of diseases version 

9 (ICD-9CM, 1980). The hospital discharge database is patient oriented, but if a 

patient is transferred from one hospital to another hospital he will be reported 

twice. Some variables make it possible to filter out patients that were treated in 

another hospital afterwards, however this is not a 100% check. 

The data are published on an annual basis. Due to some late transmittals of 

hospitals, there is a delay of 6 months between the year of discharge and the 

year in which the data is published. 

For the linking procedure data from 1997 up to 2004 were used in order to 

prepare a database with recorded patients who were admitted to a hospital in 

the years 1997-2003, so discharges in 2004 from casualties of accidents in 

2003 were added. Similarly discharges in 1997 from accidents in 1996 were 

excluded from linking. 

To limit the number of records in the hospital database to a reasonable amount, 

the linking procedure was carried out for each year separately. Fatalities and 

patients treated shorter than one day (day-treatment, opposed to clinical 

treatment) were included, in order to prevent avoidable mismatches. Records 

with an indication that it was the second or third treatment as a consequence of 

only one accident were removed, as well as records with an indication that the 

patient was treated in another hospital before. This excluded 6773 records from 

the linking process (3,3%). When applying this limitation the database contains 

200.766 records. 

Table 70: Number of patients in hospital database (admittance). 

E-code 

Description 

number of 

admittances 

1997-2003 

E810-819, excluding 817 Motor vehicle accidents 87.382 

E826-829 excluding 828 Accidents without motor vehicles 40.479 

E817+E828 Not a moving vehicle 7.404 

E820-825 Not on a public road 3.009 

E800-807 Train accident 249 

E958 Suicide 1.945 

E928+E988 Not specified 60.298 

SUM 200.766 

All these records were used in the linking process, however afterwards, records 

were excluded when they do not meet the definition of a road traffic accident. If 

a hospital record matches with a police record, it can be argued that if the police 

and the coders at AVV think it is a traffic accident it should be included. The 




people in the hospital are assumed to be less accurate in judging whether it was 

a train accident or a road traffic accident or whether the road was public or not. 

The length of stay in the Dutch hospital file is defined as the number of nights 

stayed. Besides the regular clinical stays, there are also day-treatments, where 

normally the length of stay is 1, in rare cases 2 or 3. In order to be comparable 

with the studies in other countries we agreed to redefine the Length of Stay 

(LoS) as: 

Table 71: Definition Length of Stay. 

Dutch database 

Definition used 

type of treatment length of stay 

LoS 

in LMR 

day-treatment ..


If there is no injury present in the record, the values MAIS=0 and ISS=0 are 

returned (this can be a hospitalization for observation purpose, or an injury 

outside the range 800-959 5 ). A value 9 is returned if there is an injury, but its 

severity cannot be classified by ICDmap90. Scores 1 to 6 denote increasing 

severity, from slight to untreatable. 

As our experience with this software was rather limited, the quality and 

characteristics of the medical database itself is analysed. Relations to some 

other variables are investigated: The following issues will be dealt with: 

A. The number of diagnose codes per patient; 

B. The influence of the level of detail of the injury coding on MAIS and ISS; 

C. The MAIS scores of fatalities, compared to survivors; 

D. Finally, we assess the length of stay for cases in which MAIS equals 0 or 

9 (page 132). 

Of course, results for MAIS scores depend on the number of diagnoses that are 

available. Also the level of detail that was used in the particular database is 

important. Some hospitals do not always use the lowest possible level, or data 

providers limit the level of detail to the first 3 digits, truncating the forth and fifth 

digit. 

An example ICD-9-CM Diagnosis 

Table 72: ICD9 to AIS. 

ICD9 

code 

description 

AIS 

predot 

AIS 

severity 

893 Open wound of toe(s) 

1 

893 is a non-specific code that cannot be used 

to specify a diagnosis 

893.0 Open wound of toe(s) without complication 810600 1 

893.1 Open wound of toe(s) complicated 810600 1 

893.2 Open wound of toe(s) with tendon involvement 840802 2 

The AIS scores of these injuries are 1, except for the last one, 893.2, which 

scores 2. If the database or doctor only specifies an injury code 893, the derived 

AIS severity may be underestimated. 

In order to explore the implication of these effects in the databases on the 

derived injury severity we investigated the level of detail of the Dutch hospital 

database and simulated truncation of the 5-digit code. We also simulated the 

effect of having a main diagnosis only, compared to the Dutch maximum of 9, 

see below. 

5 ICDmap90 software is able to map ICD9 codes in the range 800-959 to AIS injury codes and 

corresponding severities, except for 905-909 (late effects), 930-939 (foreign bodies) and 958 

(early complications)). So patients having one or more codes in these exclusion ranges or 

above 960 may have an underestimate of their MAIS and ISS score. In the Dutch hospital file, 

the percentage of codes that cannot be mapped by ICDmap90 is 0,53% (main diagnosis only, 

average 1984-2005), so this is only a small problem. 




A Number of injury diagnoses 

First we have examined the number of diagnoses over time. 

We analysed the 1984-2005 discharge databases, for patients that have an 

external cause with E-codes E800-E829. This selection resulted in 758.733 

injury codes (unique codes per patient). The annual number of unique injuries 

has decreased from 38.600 in 1984 to 30.900 in 2005. In the same period the 

number of patients also decreased from 22.000 to 19.400 and thus the average 

number of injuries decreases from 1,78 to 1,56 per patient. 

When we look at the distribution of the number of injury diagnoses per person, 

we see an increasing number of patients that only has one diagnose. In the last 

two decades, this has increased from 58 to 66%. The number of patients having 

no injury code at all and the number of observations (V714) are increasing and 

treated separately in Table 73. 

The observed (coded) number of injuries is different from Rosman et al. (1996), 

who used data from traffic accidents in Australia (1988). She found 10% of no 

diagnose/ observation, where we find only 1% in the same year and 5% for the 

current years of interest. The fraction having one main diagnose only is 42% 

with Rosman, 2 diagnoses 22% and all other fractions almost double the 

percentages found in the Netherlands. So compared to the Dutch database the 

number of patients without injury is low, and the number of patients having 

multiple injuries is also low. This may point to an incomplete reporting, however 

this can also be related to the number of coded diseases (diagnoses outside the 

range 800.00-999.99) or to the number of duplicate injury codes that enter the 

database. In the current analysis these are left out. In the 2003 database, the 

average number of unique diagnoses per patient is 2,77. This consists of 0,30 

diseases, 1,62 injuries and 1,08 E-codes, among which 0,25 duplicate codes. 

Table 73. Development of the number of Diagnoses. 

LMR1984-2005, filtered to patients having an E-codes in E800-E829. 

Nr of 

diagnoses 

1984 

-85 

1986 

-87 

1988 

-89 

1990 

-91 

1992 

-93 

1994 

-95 

1996 

-97 

1998 

-99 

2000 

-01 

2002 

-03 

2004 

-05 

no injury diagn 0,6% 0,6% 0,7% 0,6% 0,6% 0,8% 0,9% 0,9% 1,2% 1,1% 1,3% 

Observation 0,3% 0,2% 0,2% 0,3% 0,4% 0,5% 0,8% 2,1% 3,4% 3,7% 4,1% 

1 =main 58% 59% 59% 58% 59% 60% 61% 63% 63% 64% 66% 

2= 1 sub 22% 22% 22% 22% 22% 21% 21% 20% 19% 19% 17% 

3 10% 10% 10% 10% 9% 9% 9% 8% 7% 7% 6% 

4 4,5% 4,5% 4,5% 4,2% 4,3% 4,1% 3,8% 3,3% 3,2% 2,7% 2,4% 

5 2,2% 2,0% 2,1% 2,0% 1,9% 2,1% 1,8% 1,6% 1,5% 1,5% 1,2% 

6 1,1% 1,1% 1,1% 1,1% 1,1% 1,0% 1,0% 0,8% 0,8% 0,8% 0,7% 

7 0,6% 0,6% 0,5% 0,7% 0,6% 0,6% 0,6% 0,5% 0,5% 0,4% 0,4% 

8 0,3% 0,4% 0,3% 0,4% 0,4% 0,3% 0,2% 0,2% 0,2% 0,2% 0,2% 

9 0,3% 0,2% 0,2% 0,3% 0,2% 0,2% 0,2% 0,1% 0,2% 0,1% 0,1% 




Figure 18: Distribution of patients with their number of diagnoses 

for readability reasons the remaining 50% is left out of this figure. 

50% 

>6 6 5 4 3 2 Observation no injury diagnose 1 

40% 

30% 

20% 

10% 

0% 

1984- 

85 

1986- 

87 

1988- 

89 

1990- 

91 

1992- 

93 

1994- 

95 

1996- 

97 

1998- 

99 

2000- 

01 

2002- 

03 

2004- 

05 

From this data it can not be judged weather multiple injuries are less frequent in 

reality or that this a reporting issue, reflecting the quality of the database. 

B Level of Detail of the recorded injury codes 

When studying the level of detail of the ICD9-cm injury codes, we see that only 

a small minority of injuries is not coded at the lowest level available. Since 1992 

all injuries are coded at the lowest available level. 

Table 74: Percentage of injuries that is not coded at the lowest level. 

LMR1984-2005, filtered to E-codes E800-E829. 

% not at 

lowest level 

1984 

-85 

1986 

-87 

1988 

-89 

1990 

-91 

1992 

-93 

1994 

-95 

1996 

-97 

1998 

-99 

2000 

-01 

2002 

-03 

2004 

-05 

main diagnosis 1,08 0,78 0,52 0,17 0 0 0 0 0 0 0 

All 9 diagnoses 1,03 0,67 0,42 0,13 0 0 0 0 0 0 0 

However, some codes that are specified with all 5 digits, are not very specific. 

Compare for example the following codes: 




ICD9 AIS description 

800.00 2 Fracture of vault of skull 

Closed without mention of intracranial injury 

unspecified state of consciousness). 


Closed without mention of intracranial injury 

with prolonged [more than 24 hours] loss of consciousness and return to 

pre-existing conscious level) 


Closed with other and unspecified intracranial hemorrhage 

unspecified state of consciousness) 

It is relevant for the derived AIS score that the coded ICD9 code is sufficiently 

specific. Codes ending at '00' may show a less severe injury than a fully 

specified injury. Not in all cases it is as dramatic as in the example above, 

however its influence on derived severities should be considered as an 

explaining factor when using this kind of data and comparing across different 

hospitals, databases or countries. 

In the table below, the structure of the injury code is specified with its 

occurrence in the data. The development over time is small; the four-digit codes 

ending at "0" decreased from 14% to 10% and the 5 digit code ending on "0x" 

increased from 14% to 18%. In total for about 30% of the injury codes a more 

specific ICD-code exists with probably a more accurate severity score. 

Table 75: Number of injuries by structure of the injury code. 

LMR1984-2005, filtered to patients with E-codes E800-E829. 

Digits Structure Number of codes Distribution 

3 xxx 8.873 1,2% 

4 xxx.x 245.016 32,3% 

xxx.0 91.530 12,1% 

5 xxx.xx 133.456 17,6% 

xxx.0x 124.273 16,4% 

xxx.x0 68.035 9,0% 

xxx.00 87.550 11,5% 

Total 758.733 100% 

Apart from coding the injuries at an aggregated –non specific– level, the 

hospitals that provide data to road safety research may truncate the codes or 

only provide the main diagnose. In order to estimate the influence of this on the 

derived severity (MAIS or ISS) this is simulated further in this section. 

B1 Influence on MAIS 

In this section we simulate the influence of omitting detail of the ICD9 injury 

code and omitting subdiagnoses. We analysed the 2005 discharge database, 

that contains 31.176 records for E-codes E800-E829+E928+E958+E988. 




The most correct derivation of MAIS scores is of course found when using all 

injuries available in the most detailed coding (5 digits). 

Table 76: Effects of truncation and limitation of injury codes 

on MAIS distribution. LMR2005, N=31.176. 

MAIS 

% Truncation Main diagnosis only 

All (9) All (9) All (9) Main Main 

diagnoses diagnoses diagnoses diagnose diagnose 

5 digits 4 digits 3 digits 5 digits 4 digits 

Main 

diagnose 

3 digits 

0 no injury 6,4 6,5 16,6 6,6 7,0 19,4 

9 undetermined 5,2 5,7 5,1 6,1 6,7 5,9 

1 minor 20,4 22,9 26,8 21,7 24,8 26,7 

2 moderate 48,1 45,9 34,4 48,2 45,4 33,6 

3 severe 17,2 17,6 16,5 15,5 15,2 14,1 

4 serious 1,7 1,3 0,49 1,2 0,98 0,30 

5 critical 0,94 0,02 0,72 0,02 

6 not survivable 0,04 0,00 0,03 0,00 

Sum 100% 100% 100% 100% 100% 100% 

Comparing the share of patients of one MAIS group with the 'most correct' 

group (9 diagnoses 5 digits), we can see that the effect of truncation to 4 or 

even 3 digits removes all MAIS 5 or 6 cases (100% of the cases disappear). 

The effect of just using the main diagnosis is not too large. This is of course 

related to the fact that 70% of the patients only have the main diagnosis. Using 

only the main diagnose, truncated to 3 digits will let the number of MAIS=0 

cases increase to 3 times its value, so will also the percentage (+200%, see 

Figure 19). 

Figure 19: Effects of truncation and limitation of injury codes, 

differences with the most correct MAIS scores (all 9 codes, 5 digits). 

All (9) Diagnoses 4 digits 


Main diagnose 5 digits 



250% 

200% 

150% 

100% 

50% 

6 

5 

4 

3 

2 

1 

9 

0 

0% 

MaxAIS 

-50% 

-100% 




B2 Injury Severity Scale (ISS) 

On the same basis of AIS scores, the ISS can also be calculated. The Injury 

Severity Score (ISS) is an anatomical scoring system that provides an overall 

score for patients with multiple injuries. Each injury is assigned an Abbreviated 

Injury Scale (AIS) score, allocated to one of six body regions (Head, Face, 

Chest, Abdomen, Extremities (including Pelvis), External). Only the highest AIS 

score in each body region is used. The 3 most severely injured body regions 

have their score squared and added together to produce the ISS score. 

The ISS is comparable to the New Injury Severity Score (NISS), in which the 

limitation to body region is cancelled; just the three most severe injuries are 

taken. 

An example of the ISS calculation is shown below: 

Table 77: Calculation of ISS from AIS scores. 

Example from www.trauma.org 

Body Region Injury Description AIS Square Top Three 

Head & Neck Cerebral Contusion 3 9 

Face No Injury 0 

Chest Flail Chest 4 16 

Abdomen 

Minor Contusion of Liver 

Complex Rupture Spleen 

Extremity Fractured femur 3 

External No Injury 0 

Injury Severity Score 50 

The ISS takes values from 0 to 75. If an injury is assigned an AIS of 6 (not 

survivable injury), the ISS score is automatically assigned to 75. 

We analysed the same data by ISS, resulting in the following shifts over the ISS 

groups. For presentation we grouped ISS scores together at square values. 

2 

5 

25 




Table 78: Effects of truncation and limitation of injury codes 

on ISS distribution. LMR2005, N=31.176. 

% Truncation Main diagnosis only 

All (9) All (9) All (9) Main Main 

diagnoses 5 diagnoses 4 diagnoses 3 diagnose diagnose 

ISS 

digits digits digits 5 digits 4 digits 

Main 

diagnose 

3 digits 

0 6,4% 6,5% 16,6% 6,6% 7,0% 19,4% 

99 5,2% 5,7% 5,1% 6,1% 6,7% 5,9% 

1-3 20,4% 22,9% 26,8% 21,7% 24,8% 26,7% 

4-8 47,4% 45,3% 34,1% 48,2% 45,4% 33,6% 

9-15 16,7% 17,0% 15,8% 15,5% 15,2% 14,1% 

16-24 2,5% 2,3% 1,4% 1,2% 1,0% 0,3% 

25-35 1,1% 0,3% 0,2% 0,7% 0,0% 

36+ 0,2% 0,0% 0,0% 0,0% 0,0% 

Using main diagnosis only or using truncated values for the Injury codes, results 

in shifts to lower injury severities. 

Figure 20: Effects of truncation and limitation of injury codes, 

differences with the most correct ISS scores (all 9 codes, 5 digits). 






250% 

200% 

150% 

100% 

ISS 

50% 

0% 

36+ 

25-35 

16-24 

9-15 

4-8 

1-3 

99 

0 

-50% 

-100% 

Comparing the share of patients of one ISS group with the 'most correct' group 

(9 diagnoses 5 digits), we can see that the effect of truncation to 4 or even 3 

digits removes all ISS >= 25 cases (100% of the cases disappear). The effect of 

just using the main diagnosis is smaller, compared to truncation. This is of 

course related to the fact that 70% of the patients only have the main diagnosis. 

Using only the main diagnose, truncated to 3 digits will let the number of ISS=0 

cases increase to 3 times its value, so will also the percentage (+200%, see 

Figure 20). 




C Severity of fatalities, compared to survivors 

It is known that not all fatalities have untreatable or critical injury (MAIS is 5 or 

6). The distribution over the different MAIS-scores in the Dutch hospital file may 

be relevant for comparison with other databases; therefore a short analysis will 

be given here. 

All casualties in the hospital file (1997-2005) having a traffic accident E-code 

(E810-E829) were selected. 

Table 79: Fatalities by MAIS, casualties and percentage of fatalities with 

respect to all casualties per MAIS severity. LMR traffic 1997-2005. 

Fatalities 

Fatality distribution 

by MAIS 

MAIS 

within 30 after 30 within 30 after 30 

days days days days 

Not killed % fatalities 

0 124 7 4,7% 3,0% 8.340 1,5% 

9 74 2 2,8% 0,9% 3.714 2,0% 

1 60 6 2,3% 2,6% 26.321 0,3% 

2 196 33 7,4% 14% 92.570 0,2% 

3 941 103 35% 44% 39.433 2,6% 

4 454 33 17% 14% 3.418 12,5% 

5 762 48 29% 20% 1.924 29,6% 

6 50 3 1,9% 1,3% 41 56,4% 

SUM 2661 235 100% 100% 175.761 1,6% 

The injury of fatalities that die after 30 days is a little lower than those dying 

within 30 days. The percentage of casualties dying increases with severity, as 

was expected. 

From a split by age, we see that the more severe MAIS codes can be found with 

younger casualties, where some elderly with MAIS1 and MAIS2 also die. The 

majority of elderly fatalities has a MAIS equal 3. 




Figure 21: Fatalities by age group and MAIS. LMR traffic 1997-2005. 

400 

350 

6 

300 

5 

250 

4 

200 

3 

150 

2 

100 

1 

50 

0 

0-4 5-9 10- 

14 

15- 

19 

20- 

24 

25- 

29 

30- 

34 

35- 

39 

40- 

44 

45- 

49 

50- 

54 

age group 

55- 

59 

60- 

64 

65- 

69 

70- 

74 

75- 

79 

80- 

84 

85- 

89 

90+ 

9 

0 

Figure 22: Morbidity (number of fatalities/total number of casualties) 

by MAIS score and age group. LMR traffic 1997-2005. 

90% 

% of fatalities by MAIS and age 

80% 

70% 

60% 

50% 

40% 

30% 

20% 

5,6 

4 

3 

2 

1 

9 

0 

10% 

0% 

0- 4 5-9 10-14 15-19 20-24 25-29 30-34 35-39 40-44 45-49 50-54 55-59 60-64 65-69 70-74 75-79 80-84 85+ 

Age 

From Figure 22 it is obvious that traffic casualties have more chances to survive 

the younger they are. It is remarkable that the percentage of survivors is so 

large for MAIS3 cases, but this may be caused by the huge numbers of 

survivors, rather than by a small number of fatalities. 




D Relation between MAIS and LoS 

For some casualties there is no information on their injury. However the length 

of hospital stay is known. In this study different measures for severity are 

assessed, so it is relevant to study the cases in which one of the measures 

cannot be determined. To be more specific: if the MAIS score cannot be 

determined (MAIS=0 or MAIS=9), the length of stay can give information on the 

severity. 

In the figure below, the relation between the Maximum AIS score and the 

average Length of Stay is presented, for 9 age groups. LMR data for 1997-2005 

have been used, omitting fatalities (at any LoS). 

Figure 23: Average Length of Stay (LoS) by age for different MAIS-levels. 

LMR1997-2005, N=155.416. 

40 

35 

0 9 1 

2 

5+6 

3 4 

Average LoS 

30 

25 

20 

15 

10 

5 

0 

0-15 15-19 20-24 24-34 35-44 45-54 55-64 65-74 75+ 

Age group 

The average length of stay for injuries that cannot be classified by ICDmap90 

(MAIS9) is a little longer than MAIS1, but shorter than MAIS2. Lack of injuries in 

the range 800-959 (MAIS0) have the same length of stay as MAIS1 patients. 

This confirms that ICDmap90 is capable to score the most severe injuries and 

that no vital information is neglected by leaving out MAIS=0 and MAIS=9. 




7.6.2.3 Variables used for the linking process 

As the databases do not contain unique identifiers, casualties and patients need 

to be linked by other variables. The following variables where used as key 

variables: 

Common 

name 

Date 

Gender 

Table 80: Key variables in the hospital and police database. 

Hospital discharge database 

Police database 

Explanation Variable Explanation Variable 

Date / hour when LMRepoch Date/hour/minute when VORepoch 

entering the hospital 

the accident happens 



database) 

database) 

Gender 


database) 

LMR_gender 

Gender 

(7.627 unknowns in the 

database = 2,3%) 

P_gender 

Accident Type of the accident E_code - 

type Ecode (ICD9) 

Severity - Police indication of P_severity 

severity 

Birth Date of birth in the 

hospital file 

LMR_birth 

P_birthdate 

Region 

Province where 

hospital is located. 


database) 

Prov_zh 

Date of birth in the 

police file (24 records 

with unknown in the 

database) 

Province where 

hospital is located 

according to the police 

(for hospital treated or 

Accident & Emergency 

treated persons) or 

unknown in cases of 

very slight injury 

(casualty not 

transferred to hospital 

or died on the spot) 

P_Prov_zh 

As mentioned above, all records of injured or killed persons from the police 

database were used. From the hospital discharge database only records of 

patients involved in an accident (road accident, suicide or unspecified accident) 

were used. 

All key variables were prepared to use comparable values. The databases were 

arranged in ascending order of the variable date. To this end we presume that a 

very high percentage of all relevant road accident casualties are in a hospital 

within 4 days after the accident. Linking is done for all pairs within the timeframe 

of -1 to +4 days of the hospital admittance. 




7.6.3 Description of linking process 

In a procedure very similar to the one performed by Polak (1997, 2000, SWOV 

(2001)), it was attempted to match the police database with the hospital 

database. 

The police database contains 324.717 road traffic accidents (average 127 per 

day), whereas the hospital database contains 200.766 hospitalisations (average 

79 per day). Linking was carried out separately for each year. By simply joining 

every record of the police database with every record of the hospital database 

that meets the condition that the date difference must be in the range -1 .. +4 

days. This means entry into hospital may be 1 day in advance of the accident or 

ultimately 4 days after. As a result, approximately 18 million links are assessed 

for each year (=365 * 127 * 5*79). Of course it is not possible to be medically 

treated before the accident actually happened, but this was done to enable 

linking of records in which recording problems have occurred. 

The similarity of linked records of both databases is calculated by the distance 

function which is explained below. The quality of the link is calculated by the 

selectivity function. The selectivity points out if one can be very sure that this 

link is unique. 

We speak about LINKING if we compare two records, one from each database. 

When we examine the pairs that are best linked (they have a lower distance 

than any other pair) we speak about MATCHING. By linking and matching we 

determined the intersection of both databases. By adding relevant estimates of 

the other cells of Table 68, an estimate can be made for the total number of 

hospitalised casualties. 

7.6.3.1 The distance function 

Since no personal ID-number, which could serve as a primary key, is recorded 

in Dutch police and hospital reports, a set of other characteristics has to be 

used for matching the respective records. These key variables (see Table 80) 

were used in the distance function. The values of the key variables are 

sometimes not correctly registered or they are missing. To quantify the similarity 

between two records of the police and hospital database, a generalised 

distance function has been defined. A very low distance close to zero indicates 

a very high probability that the person in the police database is the same person 

as the one recorded in the hospital database. If the distance is larger (because 

the key values are different in some variables) the probability that this linked 

pair refers to the same casualty is smaller. 

The distance function can be described as follows: 

Let the hospital database contain N 1 records and the police database 

contain N 2 records, and let c ik denote the value of record i on key variable k 

(k = 1,…., 6, we compare records on 6 key-variables), then the distance A ij 




between record i (i=1, …, N 1 ) in LMR and record j (j = 1,…, N 2 ) in VOR is 

defined as: 

6 

6 

∑ = 1 ∑k 

= 

A ij = a = φ δ ( c , ). 

(1) 

k ijk 1 k ik 

c 

jk 

Very generally, in the term a = φ δ c , c ) 

ijk 

k 

( 

ik jk 

⎧ 

0 if cik 

= c 

jk 

⎫ 

⎪ 

⎪ 

δ ( cik 

, c 

jk 

) = ⎨ 1 if cik 

≠ cik 

⎬ (2) 

⎪ 

⎩intermediate if cik 

and/or c 

jk 

is missing⎪ 

⎭ 

and φ 

k 

a weight factor for variable k. Although the values of c ik and c jk in (2) 

are defined for each key variable, they all have in common that they 

increase the distance between two records when the records contain 

unequal categories and/or missing information on a key variable. 

The values of φ 

k 

δ ( cik 

, c 

jk 

) for the following key variables were determined on 

the assumption that a distance of 100 corresponds to a probability of about 50% 

that two records refer to the same person. 

We now continue to list the distances assigned to the key variables from Table 

80. 

1. Epoch-difference (the difference between accident and hospital entry (date/ 

time) 

aij 

= 100 * (αi – β j ) 2 /16 if α i ≥ β j ; 

a = 100 * (αi – β j ) 2 if α i < β j ; 

ij 

Figure 24: Dependence of distance to the difference in time. 

100 

80 

distance 

60 

40 

20 

0 

-1 0 1 2 3 4 

Epoch difference (days) 

In which α i is the epoch of hospital entry and β i the epoch of the accident, both 

expressed in days. This distance is constructed in such a way that it equals 100 

for a time difference of -1 and +4 days. 




2. Date of birth 

a ij 

= 220 * 0 = 0 if all 8 digits are equal; 

a ij 

= 220 * 0,2 = 44 if all digits but one are equal; 

a ij 

= 220 * 0,5 = 110 if all digits but two are equal; 

a ij 

= 220 * 0,45 = 99 if the date of birth is unknown in one of the records; 

a ij 

= 220 * 1 =220 if the dates differ on more that 2 digits. 

3. Gender 

a ij 

= 90 * 0 = 0 if known and equal; 

a ij 

= 90 * 0,5 = 45 if one of both is unknown; 

a ij 

= 90 * 1 = 90 if different. 

4. Region of hospital 

a ij 

= 50 * 0 = 0 if the provinces are equal; 

a ij 

= 50 * 1 = 50 if the provinces are not equal; 

a ij 

= 50 * 1 = 50 if unknown in which province the hospital is, or if no 

hospital admittance in the police file; 

5. Accident type (E-code), only in hospital file 

= 100 * 0,9 = 90 if E-code equals 817.*, 828.*, 958.* of 988.*; 

a ij 

a ij 

a ij 

a ij 

= 100 * 0,5 = 50 if E-code equals 820.* to 825.*; 

= 100 * 0,55 = 55 if E-code equals 928.9*; 

= 100 * 0 = 0 in all other cases: 810-816, 818, 819, 826, 827, 829 

6. P_severity, only in police file, see coding from Table 69. 

= 50 * 0 = 0 if P_severity equals 0, 2, 3, 4, 5, 6, 9 or 10; 

a ij 

a ij 

a ij 

= 50 * 0,7 = 35 if P_severity equals 1 or 8; 

= 50 * 1 = 50 if P_severity equals 7. 

At first it looks strange to give a distance of 0 if P_severity is 0, 9 or 10, while 

these values indicate that the casualty was not hospitalised. However in these 

cases there is no hospital, nor a province of the hospital, known which leads to 

a distance of 50 on variable 4. 

If a pair of records has a difference on more than one variable, the distances 

are added (see Equation (1)). 

For each pair of records that has been formed by joining the two databases, the 

distance is calculated. Small distances correspond to similar and matching 

pairs. Each record in one database forms a number of pairs with records from 




the other database. The records from the other database are called neighbours 

here. 

Table 81: Example distances from a police record to two hospital records. 

Distance 

P-h1 

Distance 

P-h2 

key police hosp1 hosp2 

date 22-1-2002 23-1-2002 23-1-2002 

hour 23 2 3 

minute 35 

Epoch 37278,98 37279,08 37279,13 0,06 0,13 

Birth 23-3-1980 23-4-1980 23-3-1980 44 0 

Gender Male male male 0 0 

Region 5 5 6 0 50 

Ecode 812 813 0 0 

Severity 6=hospitalised 0 0 

Distance 44,06 50,13 

The preferred pair in this example is the one with the correct region, but an 

acceptable typing error in the date of birth. 

The goal in the remaining part of this section is to find the best fitting neighbours 

(matching) and to tell something about the quality of the match (a combination 

of distance and selectivity), compared to other neighbours. 

7.6.3.2 The selectivity function 

By calculating the distance for two records in the two databases the similarity of 

the records can be quantified. If e.g. a record in the police database finds a very 

similar record (small distance) in the hospital database, they probably belong to 

the same person. However, it is also important to check if there are other 

hospital records which also have a small distance to the initial record in the 

police database. If others are present, you do not know which one is the true 

match and the uniqueness of the initial pair can be criticised. 

The selectivity of a matched pair is the minimum of the differences in distance to 

its next best neighbour. If the distance to the next best neighbour is large, the 

selectivity is high, also the uniqueness is perfect. If the selectivity is low it is very 

unsure which pair refers to the same person. This is the case with twins having 

an accident together. However from our anonymised view we do not know if 

they are really twins, only that in the administrations of police and hospital 

records are very similar. 

7.6.3.3 The matching procedure 

The linking and matching procedure was programmed in SAS. In this section 

the method of how to find matches between the hospital database and the 

police database is described briefly. For matches there must be a high 

probability that they refer to the same person in each database. 

First the distances are calculated for each of the 18 million pairs that are joined 

annually. 




Second, the best neighbour and the next best neighbour are determined by 

using the distance function. This is done twice, once starting from the police 

records and once starting from the hospital records. Starting from a police 

record, the smallest distance to any of the hospital records is determined (within 

the timeframe -4 to +1 days) , as well as the record number, the same for the 

smallest distance but one. 

Equal distance 

If a record has a certain distance to its best neighbour and has the same 

distance to its next best neighbour, it is uncertain which record must be taken 

for the match. 

The distance that is assigned for the epoch difference between the VOR and 

LMR record is a number with a lot of decimals. As the accident time is very 

unlikely to be exactly the same this problem mainly exists with two records in 

the hospital file (only hour is available) that link to one record in the police file. In 

these cases the first record is taken. The selectivity will be zero (as the 

difference in distance to the next best neighbour is zero) and the quality of the 

match is poor. There is only a small number of these equal distances. 

Matching 

In the third step the real matching is performed. The records meeting the 

following conditions will be assigned to each other, which means that they 

belong to the same person: 




After the complete database has been matched by the rule above, the 

remainder of the database continues with: 




The remainder continues with: 







Now many records from one database have been matched with records from 

the other database. Records that have not been selected in the four rules above 

belong to the rest-files 'police not hospital' and 'hospital not police' respectively. 




7.6.3.4 Selectivity and quality of a match 

After the matching has been done as described above, the selectivity of each 

matched pair is determined. 

First the difference between the distances of a record to its best neighbour and 

to its next best neighbour are calculated. This is done for records of both 

databases. The lowest difference of a matched pair is then stored as their 

selectivity in both records of the matched pair. 

Not all matched pairs really belong to the same person. Matches at a large 

distance do differ much on the key variables, but there seems to be no other 

closer record. This may be caused by underreporting of road accidents by the 

police. In order to separate true matches from weak matches, a combined 

criterion of distance and selectivity was used. 

Table 82 summarises the relation between distance and selectivity for all 

matched records in the Dutch police and hospital databases of 1997-2003. Of 

the 200.000 records in the hospital database, 112.000 were matched with 

records in the police database, the latter consisting of 327.000 records. 

To simplify inspection of the results, in Table 82 the observed distances for the 

matched records have been divided into seven classes ranging from very 

similar (distance class 0-0.1) to very dissimilar (distance class 220+). For the 

same reason, the observed selectivity values have been classified into five 

classes ranging from very low (selectivity class 0-10) to very high (selectivity 

class 130+). 

distance 

class 

Table 82: Frequencies of distance and selectivity classes 

for matched records in Dutch police and hospital databases 

(1997-2003, excluding fatalities and day treatment). 

selectivity class 

0-10 10-30 30-80 80-130 130+ Total 

0-0.1 244 47 1.306 13.467 17.956 33.020 

0.1-35 64 26 373 3.118 4.312 7.893 

35-55 349 147 5.510 9.850 396 16.252 

55-100 1.909 1.094 5.329 2.547 581 11.460 

100-160 7.356 5.033 4.851 502 8 17.750 

160-220 15.295 5.570 1.555 3 0 22.423 

220+ 1.198 835 153 2 0 2.188 

Total 26.415 12.752 19.077 29.489 23.253 110.986 

As can be seen in Table 82, of all the distance classes the one containing the 

smallest distances (0-0.1) has the highest frequency (33,020 records). 

Moreover, 95% of the matched record pairs in this distance class have a 

selectivity of 80 or more. In the second distance class (0.1-35), the selectivity is 

larger than 80 in 94% of the cases. On the whole, larger distance classes are 

associated with lower selectivity classes. Almost all matched record pairs with 

distances larger than 100 have selectivity values smaller than 80. 




It may safely be assumed that many of the matched pairs in Table 82 do not 

actually refer to one and the same casualty. Incorrectly matched records are 

certain to arise as a result of random matching. In the ideal situation where 

police and hospital records would contain no coding errors or missing values, 

correctly matched records could be differentiated (almost) perfectly from 

incorrectly matched records, just by evaluating their distances and selectivity 

values. Correctly matched records would all be characterised by near zero 

distances combined with large selectivity values, while incorrectly matched 

records would all be associated with large distances combined with small 

selectivity values. In this ideal situation, only the upper right and lower left cells 

of Table 82 would contain nonzero frequencies (representing the correctly and 

incorrectly matched record, respectively), while the frequencies in the upper left 

and lower right cells would all be equal to zero. Incorrect matches would for 

example arise when patients from other than traffic accidents would match to 

slightly injured casualties from the police database that did not attend hospital. 

As Table 82 shows, in reality the distinction is not so clear-cut, although the 

larger nonzero frequencies in Table 82 do tend to be concentrated in the upper 

right and lower left cells of the table. Therefore, in the next section a method is 

presented to differentiate the correctly matched records in Table 82 from the 

incorrectly matched records. 

Table 83: Matching quality status of cells in Table 82 

1 (high quality) up to 6 (uncertain quality). 

distance 

class 

selectivity class 

0-10 10-30 30-80 80-130 130+ 

0-0.1 6 6 1 1 1 

0.1-35 6 6 2 2 2 

35-55 6 6 3 3 3 

55-100 6 6 4 4 4 

100-160 6 6 5 5 5 

160-220 6 6 6 6 6 

220+ 6 6 6 6 6 

In this method, matched records are assigned to one of six matching quality 

classes. The matching quality classes are defined in Table 83. Records of 

matching quality class 1 have a high probability of having been correctly 

matched. At the same time, the correctness of the matching of records assigned 

to class 6 ranges from uncertain to improbable. In the next section, this 

classification of matched records is used to obtain estimates of the number of 

correctly matched records, and thus of the frequency in the first cell of Table 68 

"in both databases" (intersection of police and hospital file). Matches of poor 

quality, as well as records that are not matched at all, will appear in the cells 

"Only in hospital database" and "Only in police database" respectively. 

In the practical application we used matches of quality 1, 2 and 3 as part of the 

intersection of both databases, whereas we omitted matches with quality 4 5 

and 6. An exception was made for matches between police records and hospital 




records that had an Ecode of a non-traffic accident, so quality 4 is included for 

Ecodes 817+ 828+ 928+ 958+ 988). The underlying idea is that if the police say 

it is a traffic accident this information is assumed to be true, so if the hospital file 

says that it was a suicide attempt, the hospital information is overruled by the 

police information. This split of the quality 4 matches is indicated by quality 4*. 

This brings us to a total of 63.779 matches (57% of all matches). The pour 

quality matches will be added to the rest-files 'police not hospital' and 'hospital 

not police' respectively. 

7.6.3.5 Plausibility check of the matched records 

As no set of records with proofed quality of linking is available, the linking 

procedure can not be validated perfectly. To see if the linking procedure is 

calculating reasonable results, an alternative way of checking the plausibility of 

the matched records was chosen. Although available in both databases, the 

mode of transport is not one of the key variables. This gives the opportunity to 

see if the mode of transport corresponds within a matched pair. 

Table 84: Distribution of modes of transport in matches of quality 1-3+4* 

for accidents involving motorised vehicles. 89% of the cases have the 

same mode (omitting unspecified modes of transport). 

1997-2003 

Mode in VOR 

Pedes 

trian 

Bicycle Moped 

Motor 

Mode hospital file LMR 

cycle 

Car Lorry/ 

truck Other 

Not 

specified 

Total 

Pedestrian 3.058 236 31 8 105 89 11 197 3.735 

Bicycle 913 7.156 117 22 312 38 7 449 9.014 

Moped 150 350 8.543 581 187 8 280 439 10.538 

Motorcycle 31 12 236 3.884 75 1 10 166 4.415 

Car 624 226 91 68 20.820 142 19 2.063 24.053 

Lorry/truck 23 5 1 1 100 242 2 33 407 

Other 10 9 8 5 46 2 36 17 133 

Total 4.809 7.994 9.027 4.569 21.645 522 365 3.364 52.295 




Table 85: Distribution of modes of transport in matches of quality 1-3+4* 

for accidents not involving motorised vehicles. 87% of the cases have the 


1997-2003 

Mode in 

VOR 

Pedes 

trian 

Bicycle Moped 

Motor 


cycle 

Car 

Lorry/ 

truck Other 

Not 

specified Total 

Pedestrian 209 52 1 0 0 0 4 8 274 

Bicycle 53 3.097 6 0 6 0 4 33 3.199 

Moped 7 292 100 0 3 0 5 8 415 

Motorcycle 0 6 0 2 0 0 0 0 8 

Car 3 51 2 0 8 0 2 2 68 

Lorry/truck 0 2 0 0 0 2 0 0 4 

Other 1 5 0 0 0 1 11 7 25 

Total 273 3.505 109 2 17 3 26 58 3.993 

Many of the off-diagonal numbers can be explained by the confusion of 

transport mode in cases of bicyclists walking with the bike or drivers/passengers 

of motor vehicles being hit when getting off or standing next to their vehicle. If in 

the hospital file the mode is unspecified, we see that the numbers follow the 

distribution of the specified modes. 

When we compare these results for the poor quality matches (4*+5+6), we 

observe that these matrices are much more randomly distributed than the high 

quality matrices. 

Table 86: Distribution of modes of transport in matches of quality 4*+5+6 

for accidents involving motorised vehicles. 41% of the cases have the 


1997-2003 

Mode in VOR 

Pedes 

trian 

Bicycle Moped 

Motor 


cycle 

Car 

Lorry/ 

truck Other 

Not 

specified 

Total 

Pedestrian 395 104 118 67 199 22 17 95 1.017 

Bicycle 407 810 770 367 894 56 57 269 3.630 

Moped 309 346 1.829 438 877 38 125 307 4.269 

Motorcycle 79 49 167 400 250 21 17 70 1.053 

Car 726 646 1.453 963 3.654 166 160 685 8.453 

Lorry/truck 16 13 40 32 65 20 6 24 216 

Other 7 11 21 15 34 2 3 6 99 

Total 1.939 1.979 4.398 2.282 5.973 325 385 1.456 18.737 




Table 87: Distribution of modes of transport in matches of quality 4*+5+6 

for accidents not involving motorised vehicles. 26% of the cases have the 


1997-2003 

Mode in VOR 

Pedes 

trian 

Bicycle Moped 

Motor 


cycle 

Car 

Lorry/ 

truck Other 

Not 

specified 

Total 

Pedestrian 53 815 8 1 1 0 8 7 893 

Bicycle 89 4.182 20 1 14 4 32 47 4.389 

Moped 68 2.949 37 2 12 2 24 29 3.123 

Motorcycle 8 672 5 1 3 1 6 13 709 

Car 140 7.071 21 2 33 4 55 72 7.398 

Lorry/truck 6 195 0 0 1 0 3 1 206 

Other 3 73 1 0 0 0 1 0 78 

Total 367 15.957 92 7 64 11 129 169 16.796 

This confirms that matches of quality 1-4* can be considered as correct. 

Other options for a quality check, such as 

o linking the police records from one year to the hospital records from a 

different year (this should only give a small number of quality matches), 

o linking with other techniques, 

were not yet performed. 

7.6.4 Results and Discussion 

In the previous sections the intersection between the two databases was 

determined. Now we return to Table 68 in order to fill the upper left cell. 

Additional details in this table will be added, as can be seen in Table 88. 

The databases contain details of road traffic fatalities and casualties (severe as 

well as minor injuries). As fatalities and minor injury records do by definition not 

belong to the 'hospitalised road traffic casualties' their number is only included 

in Table 88 (labelled NotHRTC) and excluded in almost all other tables and 

further analyses. 

Casualties in the hospital database were not always hospitalised: some of them 

received day-treatment, whereas the majority stayed overnight or more nights. 

The ones that stayed one or more nights are hospitalised according to the 

Dutch definition. However, if we want to compare internationally, we need to 

explore common definitions of a ‘severe road traffic casualty’. Different types 

and criteria will be addressed here, based on the Length of Stay and the 

Maximum AIS. 

We will split the numbers NotHRTC apart from our basic Table 68, in order to 

omit these in the remainder of the report. This provides five groups of casualty 

records: 




1. police and hospital – casualties in police and hospital records; 

2. hospital, not police – casualties in hospital records but not in police records; 

3. hospital, not a traffic accident – casualties in hospital records that have been 

used in the linking process, but were excluded from the label 'hospitalised 

road traffic casualty' because the casualty died within 30 days, received only 

day-treatment, was not a traffic casualty (but other or unspecified external 

cause); 

4. police, not hospital – casualties in police records but not in hospital records; 

5. police not hospitalised – casualties in police records that haven been used in 

the linking process but were excluded from the label 'hospitalised road traffic 

casualty' because the casualty died within 30 days or had only minor injury. 

Groups 1, 2 and 4 also contain records that do not fulfil the definition of a 

hospitalised road traffic casualty. In group 1 (intersection) 2015 cases concern 

fatalities 6 , 1532 day-treatment. Group 2 contains 513 fatalities and 2957 cases 

of day-treatment. 

For the fourth group of records an assumption is made, while the information 

from the police is doubted: 4% of the total number of hospitalised casualties 

(79.984 according to the police) is assumed to follow the definition 'hospitalized, 

reported by police but not recognized in hospital'. All others (total – 4% - linked 

– fatalities - day treatments = 27.069 cases) are assumed not to reflect a 

hospitalisation. Reasons can be that they have the wrong severity assigned (not 

hospitalised) or that some information on key variables is incorrect in either one 

of the databases, so that a match is prevented. By taking the maximum number 

of records on the hospital database (group 2), these cases are caught without 

the risk of double counting. 

The results of the linkage can be summarised as follows: 

Table 88: Linking between 324.717 police records and 200.066 hospital 

records (1997-2003). 

1 police and hospital 

(60.232 casualties 

+ 3.547 NotHRTC) 

4 police, not hospital 


+ 27.069 NotHRTC) 

5 Police Not HRTC 

(230.664 casualties) 

2 hospital, not police 


+ 3.470 NotHRTC) 

6 neither police nor 

hospital 

(estimated 2.826) 

3 Hospitalised 

NotHRTC 

(70.163 casualties) 

The cells within the double lines form the total number of hospitalised road 

traffic casualties (129.617 casualties). 

A 6 th group of neither police nor hospital can be seen inside the double lines. 

With a Capture-Recapture methodology an estimate has been made of the 

6 2015 fatalities, of which 1779 dead in both files, 114 in police file only and 122 in medical file only. 




number of not reported casualties that fulfil the definition of hospitalised. We 

estimated the number of ‘neither police nor hospital’ casualties (2826) from the 

assumption that the following ratios are equal 

neither police nor hospital 

police not hospital 

hospital not police 

= (3) 

police and hospital 

For the real numerical application the reader is referred to Reurings et al. (2007) 

as some records (5437) from group 2, which were linked with pour quality, have 

been assigned as true links by the so-called Footprint method. The number has 

been split over the modes of transport by the same distribution as in the police 

file. 

Other national studies have not included such estimates, however they take into 

account all police reported records where we made an assumption for ‘police 

not hospital’ (group 4, see above). We may only apply this assumption if we 

also correct for not reported cases. 

We will see later that depending on a boundary value set by MAIS or Length of 

Stay only a part of these casualties will follow the new definition of Severely 

Injured. 

The results of the linkage are summarised by road user type and casualty 

severity in Table 89, while the national totals from police reported casualties are 

shown in Table 90. 

Table 89: Linkage results by road user type (1997-2003). 

hospital 

not 

hospital 

police not police 

hospitalised slight 

Total 

Car/van occupant 21.176 4.590 12.570 38.336 

Motorcyclist 3.847 831 3.266 7.944 

Moped 9.230 2.477 11.146 22.853 

Pedal cyclist 10.323 2.732 32.006 45.061 

Pedestrian 3.620 781 3.693 8.094 

Other 539 86 672 1.297 

Subtotal 48.735 11.497 63.354 123.586 

Car/van occupant 1.479 1.309 2.788 


Moped 599 530 1.129 

Pedal cyclist 644 570 1.214 


Other 48 31 79 

Subtotal 3.205 0 2.826 6.031 

Total 51.940 11.497 66.180 129.617 




Table 90: National reported totals, 1997-2003, by road user type 

and police severity. 

Road user hospitalised slightly injured Total 

Car/van occupant 36.972 107.339 144.311 

Motorcyclist 5.539 10.595 16.134 

Mopedist 14.992 50.936 65.928 

Pedal cyclist 16.048 53.199 69.247 

Pedestrian 5.326 11.577 16.903 

Other 1.107 3.678 4.785 

All 79.984 237.324 317.308 

In The Netherlands we are used to express a reporting rate as 79.984/129.617 

= 61,7%. This is similar to a general factor of 1,62, working on the number of 

hospitalised only. However in this international approach we want to establish a 

set of factors based on hospitalised and slight injuries as well. Furthermore we 

want to deselect some of the 129.617 casualties as they do not all fulfil severity 

conditions as Length Of Stay≥boundary LOS and MAIS≥boundary MAIS . 

7.6.4.1 Results for Length of Stay 

First the data are analysed by the Length of Stay (LoS) in hospital. The overall 

results of the linkage are shown in Table 91, omitting fatalities from the police 

database. The proportion of casualties who were not reported by the police is 

lower among the more severely injured, but the difference is less than has been 

found in other countries. The Length of Stay is reported for all hospital records. 

Table 91: The linkage results, by Length of Stay. 

Note that only the marked cells are considered 

as 'hospitalized road traffic casualty' 

N 

police 

not Total % not 

Length Of Stay police 'severity'= police 

reported 

(LoS) hospitalised slight 

by police 

overnight 4.919 2.140 8.761 15.820 55% 

1 10.522 3.802 15.649 29.973 52% 

hospital 

2 5.128 1.318 6.753 13.199 51% 

3 3.402 733 4.157 8.292 50% 

≥4 24.764 3.504 28.033 56.301 50% 

Fatalities within 30 days 118 4 122 

Day treatment 857 675 2.956 4.488 

not Police not in hospital file 3.205 2.826 6.031 

hospital Not hospitalised 27.069 225.148 252.217 

Total 79.984 237.324 69.136 386.444 




In order to create a table in which all casualties are distributed over Length of 

Stay and Police severity, we need to redistribute two groups of casualties over 

unknown properties. 

First we distribute 'Hospital, not police' over the severities that the police most 

probable would have assigned. We assume that for a casualty with a certain 

LoS the same proportion would be labelled hospitalised as the records that 

could be matched. From 8761 overnight casualties in the hospital database we 

estimate that a proportion of 4919/(4919+2140)=70% would be judged 

hospitalised by the police (6105 cases), and so on. 

Secondly, for the ‘police, not hospital’ casualties (3205) and for the casualties 

reported 'in neither database' (2826), no Length of Stay information is available. 

As we excluded most not linked police casualties from our results, we assume 

that these 6031 casualties have the same distribution as any other hospitalised 

casualty, so we assume that a proportion of 15820/(15820+29973+13199+ 

8292+56301)=13% stays overnight (772 cases). So the total number of 

casualties, rated hospitalised by the police that stays overnight is 4919+6105+ 

772=11796. Table 92 presents the results of both redistributions. 

Table 92: Estimated results, by Length of Stay. 

Casualties by police severity Factors Cumulative factors 


Stay hospitalised slight Total hospitalised slight hospitalised slight 

Overnight 11.796 4.796 16.592 0,147 0,020 1,322 0,101 

1 23.480 7.956 31.436 0,294 0,034 1,174 0,080 

2 11.144 2.699 13.843 0,139 0,011 0,881 0,047 

3 7.227 1.470 8.697 0,090 0,006 0,741 0,036 

≥4 52.070 6.979 59.049 0,651 0,029 0,651 0,029 

Total 105.717 23.899 129.617 1,322 0,101 

The results show that, corresponding to each casualty reported as hospitalised 

by the police (see Table 90), 23480/79984=0.294 casualties were in hospital for 

1 day, and 0.651 casualties for 4 or more days. Such conversion factors can be 

used to estimate casualty totals from police casualty totals, although changes 

over time in hospital procedures may mean that the factors depends upon the 

period chosen. 

For example, if serious casualties were to be defined as those staying 3 or more 

days in hospital then the actual total could be estimated as: 

N LoS3+ = 0,741 x number of hospitalised casualties reported by the police + 

0,036 x number of slight casualties reported by the police 

Matrix 1 by Length of Stay 

We have now determined correction factors for police records to the real 

number of hospitalised casualties for different lower boundaries of 

LengthOfStay. Now we want to do the same per road user type (mode). Table 




93 shows the data where police severity and length of stay in hospital are 

crossed per mode. 

Table 93: matrix 1, numbers by mode, police severity and length of stay. 

N 

Hospital 

not 

hospital 

Police not police 

Length police "severity"= 

Mode Of Stay hospitalised slight unknown SUM 

car/van overnight 2.971 1.174 2.298 6.443 

1 5.471 1.786 3.489 10.746 

2 2.378 529 1.327 4.234 

3 1.430 236 730 2.396 

>3 8.926 865 4.727 14.518 

motorcycle overnight 261 120 369 750 

1 586 219 825 1.630 

2 364 90 391 845 

3 307 66 240 613 

>3 2.329 336 1.442 4.107 

moped overnight 790 384 1.565 2.739 

1 1.571 718 2.650 4.939 

2 881 270 1.282 2.433 

3 651 194 793 1.638 

>3 5.337 911 4.857 11.105 

pedal cycle overnight 603 343 3.966 4.912 

1 2.029 815 7.618 10.462 

2 1.095 333 3.292 4.720 

3 754 186 2.124 3.064 

>3 5.842 1.055 15.005 21.902 

pedestrian overnight 255 111 474 840 

1 696 227 899 1.822 

2 340 79 392 811 

3 223 46 231 500 

>3 2.106 318 1.697 4.121 

other overnight 39 8 89 136 

1 169 37 169 375 

2 70 17 69 156 

3 37 5 40 82 

>3 224 19 305 548 

subtotal 48.735 11.497 63.354 123.586 

car/van unknown 1.479 1.309 2.788 

motorcycle unknown 222 198 420 

moped unknown 599 530 1.129 

pedal cycle unknown 644 570 1.214 

pedestrian unknown 213 188 401 

other unknown 48 31 79 

subtotal 3.205 2.826 6.031 

SUM 51.940 11.497 66.180 129.617 




Once more we need to use assumptions about missing information on LoS for 

records within 'police, not hospital' and 'in neither database', as well as about a 

severity that the police would have assigned for records in 'hospital, not police'. 

The same assumptions are used as above and the resulting conversion factors 

by road user type are presented in Table 94. 

Table 94: Conversion Factors based on Length of Stay and mode 

for different cut-off boundaries. 


Cumulative ≥overnight ≥1 ≥2 ≥3 ≥4 

factors hosp. slight hosp. slight hosp. slight hosp. slight hosp. slight 

car/van 0,918 0,067 0,784 0,050 0,549 0,024 0,448 0,017 0,388 0,014 

motorcycle 1,246 0,138 1,143 0,116 0,909 0,075 0,779 0,059 0,683 0,049 

moped 1,310 0,092 1,173 0,075 0,916 0,047 0,779 0,037 0,685 0,030 

pedal cycle 2,331 0,167 2,112 0,138 1,618 0,084 1,379 0,065 1,218 0,055 

pedestrian 1,314 0,129 1,195 0,107 0,922 0,067 0,791 0,053 0,709 0,046 

other 0,857 0,052 0,775 0,043 0,544 0,024 0,445 0,016 0,390 0,014 

ALL 1,322 0,101 1,174 0,080 0,881 0,047 0,741 0,036 0,651 0,029 

Further in this report these factors will be applied to the CARE data, to derive 

the number of severely injured. The bottom line in Table 94 will be referred to as 

'ALL factors'. 

Time dependency of the correction factors for LoS 

It is possible that these factors develop over time, so we split the data by year 

instead of mode and examined the annual factors. The basic annual data is not 

given here, only a graphical presentation of the resulting factors is given. We 

observe rather stable factors to be applied to hospitalised casualties. Factors for 

higher boundaries (small factors) tend to decrease. 




1,6 

1,4 

1,2 

1,0 

0,8 

0,6 

0,4 

Figure 25: Conversion factors to LoS, to be applied 

to hospitalised police reported casualties. 

Development of factors for 

hospitalized casualties 

0,2 

0,0 

1996 1997 1998 1999 2000 2001 2002 2003 2004 

≥overnight ≥1 ≥2 ≥3 ≥4 

The factor to be applied to the number of slight casualties appears to increase. 

This may be the effect of a decreasing reporting rate for slight casualties (the 

number of reported slight casualties has dropped with 27% in the 7 years 

studied), but is also influenced by the Length of Stay. For longer hospitalizations 

the factor is rather stable. If we use average factors, the results for the 

development of both factors will probably compensate. 

0,14 

0,12 

Figure 26: Conversion factors to LoS, to be applied 

to slight police reported casualties. 

Development of factors for slight casualties 

0,10 

0,08 

0,06 

0,04 

0,02 

0,00 

1996 1997 1998 1999 2000 2001 2002 2003 2004 

≥overnight ≥1 ≥2 ≥3 ≥4 




7.6.4.2 Results for Maximum AIS 

When we want to investigate the effects of a cut-off at certain MAIS level, we 

need to split the data as presented in Table 71 et seq not by Length of Stay, but 

by MAIS instead. 

The MAIS scores have been assigned from the ICD injury codes for each case. 

The overall linkage results are shown in Table 95. The column ‘% not reported 

by police’ is based on the three others, e.g. at MAIS=1 47%=8399/17846. As 

with the data in Table 90 that was based on Length of Stay, the proportion of 

casualties who were not reported by the police is lower among the more 

severely injured. 

N 

Table 95: The linkage results by MAIS. 

Police not police 

police 'severity'= 

MAIS hospitalised slight unknown Total 

% not 

reported 

by police 

0 (not known) 1.969 761 2.583 5.313 49% 

9 (unknown) 1.098 338 1.134 2.570 44% 

1 7.107 2.340 8.399 17.846 47% 

2 24.066 6.226 34.177 64.469 53% 

Hospital 

3 12.384 1.673 15.313 29.370 52% 

4 1.252 110 1.157 2.519 46% 

5 843 48 573 1.464 39% 

6 16 1 18 35 51% 

Fatality 118 4 122 

Day treatment 857 675 2.956 4.488 

Not unknown 3.205 2.826 6.031 

hospital Not hospitalised 27.069 225.148 252.217 

Total 79.984 237.324 69.136 386.444 

As with Length of Stay, in order to create a table in which all casualties are 

distributed over MAIS and Police severity, we need to distribute the two groups 

casualties over unknown properties. 

First we distribute 'Hospital not police' over the severities that the police most 

probable would have assigned. We assume that for a casualty with a certain 

MAIS the same proportion would be labelled hospitalised as the records that 

could be matched. From 8399 MAIS=1 casualties in the hospital database we 

estimate that a proportion of 7107/(7107+2340)=75% would be judged 

hospitalised by the police (6319 cases), and so on. 

Secondly, for the ‘police, not hospital’ casualties (3205) and for the casualties 

reported 'in neither database' (2826), no MAIS information is available. As we 

excluded most not linked casualties from our results, we assume that the 6031 

casualties have the same MAIS distribution as any other hospitalised casualty, 

So for example we assume a proportion of 17.846/123.586=14% to have 

MAIS=1 (871 cases). So the total number of casualties, rated hospitalised by 




the police with MAIS=1 is 7107+6319+871=14296. Table 96 presents the total 

results of this redistribution. 

Table 96: Results by MAIS and police severity. 

Casualties by police severity Factor Cumulative factor 

hospitalised slight Total hosp. slight hosp. slight 

0 (not known) 4.091 1.481 5.572 0,051 0,006 1,326 0,099 

9 (unknown) 2.090 605 2.695 0,026 0,003 1,275 0,093 

1 14.296 4.420 18.717 0,179 0,019 1,249 0,090 

2 54.364 13.250 67.615 0,680 0,056 1,070 0,072 

3 27.308 3.495 30.803 0,341 0,015 0,391 0,016 

4 2.439 203 2.642 0,030 0,001 0,049 0,0012 

5-6 1.491 81 1.572 0,019 0,0003 0,019 0,0003 

Total 106.080 23.537 129.617 1,326 0,099 

Due to small numbers the groups for MAIS=5 and MAIS=6 have been taken 

together. 

The results show that, corresponding to each hospitalised casualty reported by 

the police (see Table 90), 27.308/79.984=0,341 casualties were in hospital with 

MAIS=3, and 0,049 with MAIS 4, 5, or 6. Such conversion factors can be used 

to estimate casualty totals from police casualty totals, although changes over 

time in hospital procedures may mean that the factors depends upon the period 

chosen. 

For example, if serious casualties were to be defined as those having at least 

MAIS=2 then the actual total could be estimated as: 

N mais2+ = 1.070 x number of hospitalised casualties reported by the police + 


Comparing the redistributions for MAIS with LoS 

From 49.633 cases there is no information for the police estimate of hospitalized 

or slight injury (129.617 – 79.984). Where the Length of stay approach assigned 

25.729 of them to hospitalized (51,8%), the MAIS approach assigned 26.092 

cases to hospitalized (52,6%). So both approaches result in a comparable 

estimate. 

Matrix 2 by Maximum AIS 

We have now determined correction factors for police records to the real 

number of hospitalised casualties for different lower boundaries of Maximum 

AIS. Now we want to do the same for the road user type (mode of transport). 

Table 97 shows the data where police severity and maximum AIS are crossed 

per mode. 




Table 97: Matrix 2, numbers by mode, police severity and MAIS 

N 

Hospital 

Police 

police "severity"= 

Not 

police 

mode MaxAis hospitalised slight unknown SUM 

car/van 0 1.431 548 1.201 3.180 

9 674 212 417 1.303 

1 4.508 1.358 2.687 8.553 

2 9.147 2.007 5.515 16.669 

3 4.539 433 2.278 7.250 

4 509 20 294 823 

5 359 11 172 542 

6 9 1 7 17 

motorcycle 0 70 22 94 186 

9 61 13 44 118 

1 334 129 375 838 

2 2.088 515 2.026 4.629 

3 1.082 135 611 1.828 

4 113 11 68 192 

5 99 6 46 151 

6 2 2 

moped 0 176 89 350 615 

9 157 50 203 410 

1 867 378 1.352 2.597 

2 4.814 1.517 6.524 12.855 

3 2.786 404 2.368 5.558 

4 224 21 220 465 

5 205 18 126 349 

6 1 2 3 

pedal cycle 0 187 84 771 1.042 

9 146 49 398 593 

1 944 332 3.384 4.660 

2 5.687 1.653 17.679 25.019 

3 2.929 562 9.109 12.600 

4 300 42 471 813 

5 124 10 188 322 

6 6 6 12 

pedestrian 0 77 17 138 232 

9 50 12 56 118 

1 329 108 499 936 

2 2.081 499 2.083 4.663 

3 937 127 798 1.862 

4 93 15 89 197 

5 53 3 31 87 

6 0 

Other 0 28 1 29 58 

9 10 2 16 28 

1 125 35 102 262 

2 249 35 350 634 

3 111 12 149 272 




Not 

hospital 

4 13 1 16 30 

5 3 10 13 

6 0 

subtotal 48.735 11.497 63.354 123.586 

car/van unknown 1.479 1.309 2.788 

motorcycle unknown 222 198 420 

moped unknown 599 530 1.129 

pedal cycle unknown 644 570 1.214 

pedestrian unknown 213 188 401 

other unknown 48 31 79 

subtotal 3.205 2.826 6.031 

SUM 51.940 11.497 66.180 129.617 

Again we need to use assumptions about missing information on MAIS for 

records within 'police, not hospital' and 'in neither database' as well as about a 

severity that the police would have assigned for records in 'hospital, not police'. 

The same assumptions are used as above and the resulting conversion factors 

by road user type are presented in Table 98. 

Table 98: Conversion factors based on MAIS and mode 

for different cut-off values. 

Maximum AIS 

cumulative ≥1 ≥2 ≥3 ≥4 ≥5 

factors hosp. slight hosp. slight hosp. slight hosp. slight hosp. slight 

car/van 0,826 0,055 0,638 0,036 0,233 0,007 0,038 0,001 0,015 0,0002 

motorcycle 1,200 0,131 1,078 0,110 0,371 0,022 0,060 0,002 0,027 0,0008 

moped 1,259 0,086 1,120 0,072 0,402 0,015 0,054 0,001 0,023 0,0005 

pedal cycle 2,266 0,154 2,037 0,132 0,768 0,033 0,068 0,002 0,020 0,0004 

pedestrian 1,260 0,122 1,117 0,102 0,377 0,021 0,051 0,002 0,016 0,0004 

other 0,803 0,048 0,645 0,033 0,225 0,008 0,032 0,001 0,010 0,0001 

ALL 1,249 0,090 1,070 0,072 0,391 0,016 0,049 0,001 0,019 0,0003 

Further in this report these factors will be applied to the CARE data, to derive 

the number of severely injured. The bottom line in Table 98 will be referred to as 

'ALL factors'. 

The results emphasise that pedal cyclist casualties are recorded less fully by 

the Dutch police than casualties among other road user groups, as these factors 

are much higher. 

Time dependency of the corrections factors for MAIS 

It is possible that these factors develop over time, so we split the data by year 

instead of mode and examined the annual factors. The basic annual data is not 

given here, but a graphical presentation of the resulting factors is given. 

We observe rather stable factors to be applied to hospitalised casualties. 

Factors for higher boundaries (small factors) are also stable, but very small. 




1,6 

1,4 

1,2 

1,0 

0,8 

0,6 

0,4 

0,2 

0,0 

Figure 27: Conversion factors to MAIS, to be applied 

to hospitalised police reported casualties. 


hospitalized casualties 

1996 1997 1998 1999 2000 2001 2002 2003 2004 

≥1 ≥2 ≥3 ≥4 ≥5 

Figure 28: Conversion factors to MAIS, to be applied 

to slight police reported casualties. 

0,12 

0,10 


slight casualties 

0,08 

0,06 

0,04 

0,02 

0,00 

1996 1997 1998 1999 2000 2001 2002 2003 2004 

≥1 ≥2 ≥3 ≥4 ≥5 

7.6.4.3 Overview and discussion 

The linking between the Dutch hospital database and the police database of 

traffic casualties resulted in an estimate of 18.500 hospitalised traffic casualties 

annually, distributed over the different cells of Table 99. 




Table 99: Distribution of police and hospital records over 

intersection and rest files. 

police and hospital 

46% 

police, not hospital 

2,5% 

hospital, not police 

49% 

neither police nor hospital 

2,2% 

When applying a certain cut-off at a lower boundary on LoS or MAIS, this 

distribution will not really change. So, if only the hospital file is used, 95% of the 

records consist of real information; only 4,7% is estimated. This means that the 

distribution of casualties by variables that are in the hospital database (such as 

LoS, MAIS, mode, age, day of week, time of day) can be determined quite 

accurately, close to the real distribution. 

If only the police file is used, known information accounts only for less than 

50%. Therefore, variables that are only in the police file (such as area type [built 

up area, urban area, motorway], crash opponent, intersection/junction, weather 

condition) can not be determined very accurately. 

One of the aims of this project is to develop factors on the police file, in order to 

estimate the real numbers and their distributions. With several assumptions for 

missing information, correction factors have been developed that can be applied 

to the CARE database (i.e. the police file), resulting in an estimate of the real 

number of severely injured. 

We can distinguish different groups of variables which may be expected to be 

more or less reliable in the estimate of the real number. 

1) Variables that are present in both databases. 

Some variables have equivalents in the other database. This group can be 

split into 2 subgroups: 

a) variables that were used in the linking process (date/time, age, gender, 

severity, region), 

b) variables that were used to calculate the correction factors (severity, 

mode), 

For group b, of course the outcome reflects the real distribution for that 

variable, because the factors were developed to do so. The variable severity 

(police severity, MAIS or LoS) is a special one in this group, as it is also 

used to set a minimum on the severity, by filtering out only cases above 

certain lower boundary. In the analysis of results some assumptions have 

been made on the relations between these variables. 

Variables from 'group a' which were not used in the calculation can give us a 

validation of the outcomes (year, age, gender, day of week, time of day). 

Only if the numbers/distributions obtained in such a validation are within a 

reasonable similarity with the real distribution, one may expect the 

distributions on variables that are in one file only to be reliable. In the next 

sections we will compare the real distribution (from 95% hospital data) with 




the calculated results (police file x correction factors) for the variables year 

and age. 

2) Variables that exist in one file only. 

For variables which are available in only one database it is very difficult to 

say if the observed distribution is comparable to the real distribution, as we 

do not know this real distribution. If, for example, all accidents at intersections 

were reported correctly by the police, all missing information should 

be assigned to sections. An average correction factor as is calculated here 

would take the same fraction for both types and thus overestimate the 

number of intersection accidents and underestimate the section accidents. 

The variables Year and Age are explored below. 

7.6.4.4 The application to the police file 1991-2005 

As an example we present the number of severe casualties by mode of 

transport according to different boundary levels. We used the correction factors 

by mode of transport, but without time dependence (see Table 94 and Table 

98). We need to judge whether the results seem reliable and we compare the 

sum of all modes with the application of correction factors without mode of 

transport (see Table 92 and Table 96). 

Figure 29: Car occupants by MAIS2+, MAIS3+, LoS2+, LoS3+. 

5000 

4000 

Cars and vans 

MAIS2+ 

MAIS3+ 

LoS2+ 

LoS3+ 

3000 

2000 

1000 

0 

1990 1995 2000 2005 




Figure 30: Motorcycles by MAIS2+, MAIS3+, LoS2+, LoS3+. 

1400 

1200 

1000 

800 

600 

400 

200 

Motorcycles 

MAIS2+ 

MAIS3+ 

LoS2+ 

LoS3+ 

0 

1990 1995 2000 2005 

Figure 31: Mopeds by MAIS2+, MAIS3+, LoS2+, LoS3+. 

3500 

3000 

2500 

2000 

1500 

1000 

500 

Mopeds 

MAIS2+ 

MAIS3+ 

LoS2+ 

LoS3+ 

0 

1990 1995 2000 2005 

Figure 32: Bicycles by MAIS2+, MAIS3+, LoS2+, LoS3+. 

8000 

7000 

6000 

5000 

4000 

3000 

2000 

1000 

Bicycles 

MAIS2+ 

MAIS3+ 

LoS2+ 

LoS3+ 

0 

1990 1995 2000 2005 




Figure 33: Pedestrians by MAIS2+, MAIS3+, LoS2+, LoS3+. 

1800 

1600 

1400 

1200 

1000 

Pedestrians 

MAIS2+ 

MAIS3+ 

LoS2+ 

LoS3+ 

800 

600 

400 

200 

0 

1990 1995 2000 2005 

Figure 34: Other modes by MAIS2+, MAIS3+, LoS2+, LoS3+. 

200 

150 

MAIS2+ 

MAIS3+ 

LoS2+ 

LoS3+ 

Other 

100 

50 

0 

1990 1995 2000 2005 

As the basis for each graph is the same (the number of police reported 

hospitalised and slight casualties by mode of transport) and the factors are 

constant for each year, the patterns observed within each graph are the same. 

The level is the only difference. We can see that the lower boundaries of 

MAIS2+ and MAIS3+ form the most extreme values. LoS2+ and LoS3+ are in 

between. 

We can add up the numbers by mode of transport to reach the total number of 

casualties per severity boundary, see Figure 35. 




Figure 35: Sum of modes by MAIS2+, MAIS3+, LoS2+, LoS3+. 

18000 

16000 

Sum of modes 

14000 

12000 

10000 

8000 

6000 

4000 

MAIS2+ MAIS3+ 

2000 

LoS2+ 

LoS3+ 

0 

1990 1995 2000 2005 

Then we can compare the results if we would have applied the ALL factors from 

Table 92 and Table 96. 

Figure 36: Differences between sum of modes and the application 

of ALL factors, irrespective of modes. 

1000 

800 

600 

difference All - Sum(modes) 

400 

200 

0 

-2001989 1991 1993 1995 1997 1999 2001 2003 2005 

-400 

-600 

MAIS2+ 

-800 

MAIS3+ 

-1000 

LoS2+ 

-1200 

LoS3+ 

From Figure 36 it can be seen that there is a difference between the application 

of ALL factors and application of factors for each mode separately. The 

difference is acceptable in the period that the linking was performed and on 

which basis the factors were determined (1997-2003). However in years before 

and after, the difference is larger. For factors on LoS, the application of ALL 

factors lead to higher totals than the sum of the separate modes. With MAIS, 

the application of ALL factors in general leads to lower totals than the sum of 

separate modes. 




7.6.4.5 Age distribution Hospital - Police 

In the second experiment to explore the reliability of the results obtained, we 

looked at the age distribution. The age of the casualty was not used in the 

calculation of the correction factors. There were two main reasons not to do so: 

1) If the numbers were split by age group, too few would remain to calculate 

reliable correction factors. 

2) There is evidence that the age of traffic participants is correlated to their 

mode of transport. 

From a point of mobility the use of a certain mode of transport can be 

addressed to specific age groups quite well: cycling is dominated by teenagers, 

moped riding is done by 16-20 year olds, and public transport is mainly used by 

people in their twenties. The car is dominant at every age, but the young and 

the very old often travel as a passenger. 

Figure 37: Number of travelling kilometres per person per day in The 

Netherlands, Sources AVV(MON), CSB(OVG) 1994-2006. 

kilometeres per person per day 

45 

40 

35 

30 

25 

20 

15 

Train 

Other 

BusTramMetro 

Car (passenger) 

Car (driver) 

Motorcycle 

Moped 

Bicycle 

Pedestrian 

10 

5 

0 

0 - 11 12 - 17 18 - 24 25 - 29 30 - 39 40 - 49 50 - 59 60 - 74 75+ 

Age 

However, the insights that the accident rate is so different for beginners 

compared to that of experienced drivers and that the elderly are very fragile 

when involved in an accident help to understand that age and mode of transport 

are not very much correlated in terms of traffic casualties. 

In a comparison of the age-distribution of hospitalised and slight injuries in the 

police database with the age-distribution in the hospital file by MAIS, it can be 

observed that age distributions in both files are not the same. The hospital 

records are assumed to represent the real distribution, as the real numbers are 

built for 95% from hospital records. 




Figure 38: Age distribution of police records and hospital records 

for different severities. 

6% 

5% 

4% 

3% 

Hospitalized 

Slight 

MAIS4+ 

MAIS3 

MAIS2 

MAIS1 

2% 

1% 

0% 

0 10 20 30 40 Age 50 60 70 80 90 

The age distribution of MAIS3 casualties is the most different compared to the 

other severities (much lower for 20-50 year olds, much higher above the age of 

60). The strong peak that is observed in the police records for 17 and 18 year 

olds is not seen so strongly in any MAIS distribution of the hospital file. The high 

MAIS2 level for 5-14 year olds is not observed in the police file. 

Due to the smaller number of cases the pattern for MAIS4+ is the most varying 

one. 

Table 100: Number of cases by mode and severity. 

Police and hospital files 1997-2005. 

Source severity pedestr bicycle moped motorcycle car/van other SUM 

Police Hospitalised 6457 20069 18159 6909 45035 1368 97997 

file Slight 13511 63702 58534 12724 124168 4169 276808 

MAIS4+ 433 1583 781 453 1436 500 5186 

Hospital MAIS3 2629 18000 4736 2330 7285 2416 37396 

file MAIS2 6539 35692 11984 6066 17101 4856 82238 

MAIS1 1430 6635 2441 1085 9442 1477 22510 

The difference in age distribution may be caused by different age distributions 

per mode of transport. There are various ways to show that this is not the case. 

As an example the factors between the age distributions are plotted for cyclists 

below. 




Figure 39: Factor between the age distribution of hospital records and 

hospitalised police casualties by MAIS level. LMR+VOR 1997-2005. 

7 

6 

5 

4+/hosp 

3/hosp 

2/hosp 

1/hosp 

4 

3 

2 

1 

0 

0 20 40 60 80 

Age 

Figure 40: Factor between the age distribution of hospital records and 

slight police casualties by MAIS level. LMR+VOR 1997-2005. 

7 

6 

5 

4+/slight 

3/slight 

2/slight 

1/slight 

4 

3 

2 

1 

0 

0 20 40 60 80 

Age 

This illustrates clearly that any factor on the police records will never result in an 

age distribution that is similar to the hospital file, if the age is not among the 

components used to determine the factors. Especially the numbers of casualties 

above 60 and below 10 years old are too low; for the other ages the number is 

too high. 


Linking police reported traffic casualties and hospital patients having accident 

injuries was successfully carried out. On average 46% of the hospitalised 

casualties was found in both database, 49% was in the hospital database only, 

2,5% in the police database only. It was estimated that 2,2% of casualties was 

reported in neither database. 




Severity estimates from the hospital file, expressed in MaximumAIS and Length 

Of Stay enabled to judge the severity assigned by the police. The more severe 

the injury, the better the reporting by the police. 

As traffic is becoming safer, the number of fatalities to be studied is decreasing, 

which leads to statistical interpretation problems. A small difference with the 

expected trend may not be statistically significant. Analysis on severely injured 

casualties had the problems of underreporting and not being representative, 

next to the international comparability component. These linking studies and 

severity assessments from the hospital injuries enables to define sharp 

boundaries to groups of which underreporting coefficients are well known. With 

a boundary on MAIS3+, the group of Killed and Severly Injured (KSI) will be 

approximately 6 times lager than the group of killed only. However with a 

boundary on MAIS2+ the group of about 15 times larger. In The Netherlands it 

is technically and statistically possible to define these boundaries in a reliable 

way, so a definition of MAIS2+ is recommended for The Netherlands as study 

group. Unfortunately this appeared not feasible in the international setting. 

Correction factors have been derived to calculate the real number of severely 

injured casualties (above several severity boundaries) from the police casualty 

data. 

Problems have been observed in the application of these factors as not all 

resulting distributions follow the expected distribution. An example on age was 

given to illustrate this problem. Time dependency of the factors is another 

problem that needs to be addressed in any next study. Special attention should 

be given to the reliability of the derived factors. By disaggregating the data to 

age group and year, the numbers in each cell might become too small to 

estimate the factors with the required accuracy. 

Compared to other countries, in the Netherlands we were very lucky to have 

complete national hospital files for a larger number of years (1997-2003). Other 

countries, being less fortunate with the numbers of hospitals and years, already 

experienced the small number problems right now, only using the 

disaggregation by mode of transport. 

For international comparison, problems were encountered in the definition of the 

Maximum AIS, as not all countries have the same ICD version, the same 

number of diagnoses, and the same level of detail of the available injury codes. 

Special attention in this national report was given to the quality and 

characteristics of the medical file, such as the number of diagnoses and the 

level of detail of the diagnoses. By simulation of truncation and omitting 

diagnose codes valuable reference has been set. 

The difference of severity is influenced both by the ICD version and the AIS 

version that was used. In The Netherlands ICD9-CM was used and a 

conversion to AIS1990, while other countries used ICD10 and a conversion to 




AIS1998. For an analysis of the compound result on injury severity (MAIS) see 

the study from the United Kingdom. 

These difficulties in defining a comparable severity in all European countries 

prevented the calculation of correction factors in all countries. Further research 

on medical coding of injury and possible conversions between them is required 

before any European traffic injury severity boundary can be set. 


Hook, E.B., Regal, R.R., (1995). Capture–recapture methods in epidemiology: 

methods and limitations. Epidemiol. Rev. 17 (2), 243–264. 

ICDMAP-90 User's Guide. The Johns Hopkins University & Tri analytics, Inc. 

(1998-2002). 

Polak, P.H. (1997). Registratiegraad van in ziekenhuizen opgenomen 

verkeersslachtofers; Eindrapport. R-97-15. SWOV, The Netherlands. 

Polak, P.H. (2000). De aantallen in ziekenhuizen opgenomen verkeersgewonden, 

1985 – 1997; Koppeling van gegevens van de verkeersongevallenregistratie 

en de registratie van de ziekenhuizen. R-2000-26. SWOV, The 

Netherlands. 

Reurings, M.C.B., Bos, N.M., van Kampen, L.T.B. (2007). Berekening van het 

werkelijk aantal in ziekenhuizen opgenomen verkeersgewonden 1997-2003, 

Methode en resultaten van koppeling en ophoging van bestanden. R-2007-8. 

SWOV, The Netherlands. 

Rosman, D.L., Knuiman, M.W., Ryan, G.A. (1996), An Evaluation Of Road 

Crash Injury Severity Measures, Accident Analysis & Prevention, Vol. 28, No. 2, 

pp. 163-170. 

SWOV (2001). A new linking procedure for the determination of the total 

number of hospitalised road traffic casualties by comparing police and hospital 

reports. The Netherlands. 

Wittes, J., Colton, T., Sidel, V., (1974). Capture–recapture methods for 

assessing the completeness of case ascertainment when using multiple 

information sources. J. Chronol. Dis. 27, 25–36. 




7.7 Study carried out in Spain 

Report prepared by Catherine Pérez (ASPB) 


This section describes the national study carried out in Spain to achieve the 

aims of Task 1.5 of the SafetyNet IP. The objective of the task is to estimate the 

actual numbers of casualties from the CARE database. The general 

methodology has been previously described, therefore this section presents 

only the specific details of the Spanish study. 

As a first step, we explore the feasibility of knowing whether it would be possible 

to carry out a probabilistic record linkage with the National Accident Police 

Registry (Dirección General de Tráfico, DGT) and the National Hospital 

Discharge Registry (HDR). Both registries have a national coverage. As a result 

of the first phase we selected an Autonomous Region where the study was 

feasible: Castilla y León. We describe the characteristics of this region and the 

methodology used, then present the results for this region. 

Characteristics of Castilla y Leon 

Castilla y Leon is the largest autonomous 

region in Spain, and is located in the upper 

centre of the Iberian Peninsula. It has an 

area of 94,233km 2 which represents 18,8% 

of all Spain and 2.523.020 habitants, 5,7% 

of all Spanish population. It is divided into 9 

provinces. There is one city with more than 

300.000 population, three between 100.000 

and 200.000, and five between 50.000 and 

99.999 habitants. In 2004 there were 

1.440.056 vehicles registered, 5% of all 

vehicles registered in Spain. There are 

18,890 km of roads, 11% of the total. 

Figure 41: Map of Castilla y León 

The National Traffic Authority (Dirección General de Tráfico, DGT) reported for 

year 2005, 9,857 road victims. Among them 384 were fatalities, 2,207 severe 

injured and 7,266 slight injured. In 2004 there were 2,367 hospitalisations due to 

road injuries. 

In 2004 there were 21.0 road fatalities per 100.000 males and 5,4 road fatalities 

per 100.000 females, higher than the national estimates. (National Standardised 

mortality ratios are 17.2 and 4.9 respectively). During the same year there were 

125.3 hospitalisations due to traffic injuries per 100.000 males and 43.9 per 

100.000 females. (National estimates are 118.5 and 45.5 respectively). 





Hospital Discharge Register (HDR) database 

The Hospital Discharge Register (HDR) database was provided by the 

autonomous department of health of Castilla y Leon. Data included records of 

hospitalisations due to road injuries from 1 st July 2005 to 31 st December 2005. 

Case definition of road casualty: 

We consider a road casualty if it fulfils the following criteria: 

1. Suffering a injury defined by the International Classification of Diseases, 

9 revision, Clinical Modification. (ICD-9-CM: 800 TO 959.9) and 

2. Primary or Urgent hospitalisation (in front to scheduled, which would 

indicated that the hospitalisation is related to complications or 

rehabilitation process, and not due to a recent accident) and 

a. Type of funding: road traffic insurance company or 

b. External cause of injury: road traffic accident (E-Code: 810-819, and 826) 

No severity level is provided by ICD-9-CM. A conversion from ICD-9-CM 

diagnosis to AIS was carried out using the method developed by MacKenzie 

and implemented in ICDMAP90 software. 

Police data (DGT) 

The Police (DGT) data on fatalities or injured in Castilla y Leon was provided by 

the National Traffic Authority for the year 2005. Only cases from the second half 

year were included for the linkage process. 

Representativeness of data from Castilla y Leon 

In this section we assess how representative are data from Castilla y Leon of 

road accident casualties in the rest of Spain. We use the national DGT registry 

for year 2005 and the national HDR for year 2004. Police data shows that 

among Castilla y Leon injured there are slightly more females, less children and 

youth, and more adults over 44 years old (Table 101). Regarding the vehicle, 

there are more car users and less motorcycle and moped users. 

For the area where happened the crash, there is no difference for fatalities, but 

the proportion of non urban casualties is 11% higher than the rest of Spain for 

severe casualties and there are 21% more for slight casualties. Complementary 

the proportion of urban severe and slight casualties is lower among Castilla y 

Leon crashes. 




Table 101: Characteristics of road victims in Castilla y Leon and rest of 

Spain by level of severity. DGT Spain, 2005. 

Fatal Serious Slight 

Rest 

Castilla of Castilla y Rest of Castilla y Rest of 

y Leon Spain Total Leon Spain Total Leon Spain Total 

N 384 3 473 3 857 2 207 20 237 22 444 7 266 103 684 110 950 

GENDER 

Male 76.6 79.0 78.7 69.4 72.2 71.9 61.8 62.3 62.3 

Female 23.4 20.2 20.5 30.4 26.5 26.9 37.3 35.6 35.7 

Unknown 0.9 0.8 0.3 1.2 1.1 0.9 2.1 2.0 

AGE 

75 12.2 7.3 7.8 5.9 4.5 4.6 3.5 2.4 2.5 

VEHICLE 

Car 64.3 54.3 55.3 57.1 46.8 47.8 70.1 56.8 57.7 

Motorcycle 5.2 11.2 10.6 8.5 12.6 12.2 3.2 9.4 9.0 

Moped 3.1 6.4 6.1 6.4 16.2 15.3 5.1 15.3 14.6 

Bicycle 2.3 1.8 1.8 2.1 2.0 2.0 1.3 1.6 1.6 

Bus 1.3 0.5 0.6 0.8 0.7 0.7 1.1 1.8 1.8 

Truck or lorry 10.4 8.8 8.9 11.7 7.3 7.8 10.4 6.3 6.6 

Other 2.1 2.5 2.4 2.6 1.8 1.9 1.9 1.2 1.2 

Unknown 11.2 14.6 14.2 10.8 12.4 12.3 6.9 7.5 7.5 

AREA 

Non Urban 85.2 84.7 84.7 77.5 67.2 68.2 68.1 47.2 48.6 

Urban 14.8 15.3 15.3 22.5 32.8 31.8 31.9 52.8 51.4 

The lethality (number of deaths per 1000 victims) for non urban accidents is 

similar in Castilla y Leon than the rest of Spain. But the lethality for urban 

accidents is clearly higher (Table 102). 

Table 102: Lethality (number of deaths per 1000 victims) in Castilla y Leon 

and rest of Spain. DGT, Spain 2005 

Castilla y Leon Rest of Spain Total 

Non urban 46,8 44,9 45,1 

Urban 19,9 8,6 9,1 

Total 39,0 27,3 28,1 

There is no reason to believe that severity of crashes is higher in Castilla y Leon 

than in the whole Spain. On the contrary, it suggests that there might be a 

significant underreporting of severe and especially slight casualties. Most slight 

casualties occur in urban areas, where motorcycles and mopeds are very 

popular in Spain. It could also explain the differences by type of road user. 




Among hospitalised casualties there is a slightly lower proportion of females, 

and youth from 15 to 29 years (Table 103). Regarding severity among Castilla y 

Leon records there is a higher proportion of unknown severity. 

Table 103: Characteristics of HDR road injured in Castilla y Leon and rest 

of Spain by level of severity. HDR Spain, 2004. 

Castilla y 

Leon 

Rest of 

Spain Total 

N 2 196 29 812 32 008 

GENDER 

Male 73.7 71.5 71.6 

Female 26.3 28.5 28.4 

AGE 


as reference weight) has been determined, we check whether there is enough 

information to achieve it. The reference weight is calculated with this 

expression: 

p = probability that a pair is correct 

A = size of database 1 

B = size of database 2. 

E = number of correct pairs 

W t = log 2 ( p / (1-p)) – log 2 (E / (A x B-E)) 

The value of the weight to contrast with W t will be the attainable minimum 

weight when all the variables agree exactly (Wmin). This will be obtained from 

the combination of the most frequent categories of each variable. For instance if 

we had only sex and the position in the vehicle, the minimum weight would be 

for those records in which the category of the variable sex was “man” and the 

one of the position was “driver” because they are the most frequent. Once 

determined, it will be compared with W t (Newcombe, 1988). 

If W min >= W t Sensibility or Specificity of the process > 95% 

If W t – 3


Table 105: Distribution of cases and categories among autonomous 

regions 

Period Year 2005 

Variables Gender, age, date of accident/hospitalization. 

Autonomous Cases Cases HDR + 

Feasi 

regions 

HDR a DGT a x Common Variables 

DGT 

bility 

Baleares 1.000 5.000 6.000 2 x 100 x 365=73.000 

Galicia 2.500 9.000 11.500 2 x 100 x 365=73.000 

Castilla-León 3.000 10.000 13.000 2 x 100 x 365=73.000 

C.Valenciana 4.000 13.000 17.000 2 x 100 x 365=73.000 

ª: Estimated cases. 

We assume that the injured are hospitalised in the same autonomous region where the 

accident occurred 

If we calculate the reference and minimum weights for these autonomous 

regions only Castilla y Leon and Galicia are feasible. Table 106 shows the 

results of reference weight (W t ) and the minimum weights (Wmin) for different 

autonomous regions. Those in black show cases where it is possible to reach 

values of Sensitivity and Specificity between 90% and 95%. In no case, with 

these values it is possible to reach values of Sensitivity and Specificity greater 

than 95%. Therefore we decided to carry out the national study in the 

autonomous region of Castilla y Leon. 

Table 106: Reference weights and minimum weights to assess the 

feasibility of probabilistic record linkage 

ESPAÑA 

(2.001) 

BALEARES 

GALICIA 

CAS-LEÓN 

C.VALENCI 

ANA 

W t ( p=0,95) 20,34 15,35 16,21 16,54 16,91 

Minimum Weight (Wmin). 

Gender + age + hospitalisation date 14,04 10,29 11,83 11,83 11,11 

Gender + age + hospitalisation date + 

hospital province 

Gender + age + hospitalisation date + 

hospital reference area 

* Not available 

16,44 10,43 13,33 13,95 11,25 

* 11,74 13,52 14,46 13,83 

The method used for linking records was a mix, that use probabilistic, with the 

aid of deterministic to generate blocking and with a final manual review. The 

probabilistic linkage process consists in matching two or more records which 

are believed to belong to the same individual. It is based in two probabilities: the 

probability of matching given that both records belong to the same individual 

and the probability of matching by chance. The less probable is a value of the 

variables, the greater is the weight assigned. The process is done by the 

software WCONNECTA developed by the Agència de Salut Pública de 

Barcelona (ASPB) (Cirera et al, 2000 and Arribas et al, 2004). 




The matching process implies these phases: 

1. Data preparation 

2. Selection of linkage variables 

3. Evaluation of process feasibility 

4. Computation of simple weights 

5. Restriction of comparison pairs (blocking) 

6. Comparison stage (matching) 

7. Simple weights assignment 

8. Computation of composite weights 

9. Decision stage (linking) 

10. Threshold determination 

11. Review of dubious pairs 

1. Data preparation 

During July-December of 2005, 1,636 people were admitted a public hospital 

(HDR) of Castilla y Leon as a result of the injuries suffered in a traffic accident. 

During the same period, 6,970 victims of different severity were reported by 

police (DGT). 

2. Selection of linkage variables 

Variables used for linkage were age, gender, date of the crash and province of 

accident / province of hospital. The absence of a common identifier for both 

databases resulted in the need to link both sources of information with the 

probabilistic method, using the information of the common variables in both 

databases. We assume that the hospitalisation is done the same day of the 

crash, and in the same province of the crash. 

3.Evaluation of process feasibility 

The feasibility study has been previously explained in section 7.7.1. with the aim 

of justifying why we selected the autonomous region of Castilla y Leon. 

4. Computation of simple weights 

Prior to linking records it was necessary to compute weights that will become 

useful later in the linking phase. These weights are based on two probabilities, 

the probability of matching given that both records belong to the same 

individual, and the probability of matching by chance. The less probable is a 

value of the variable, the greater is the weight assigned (Jaro, 1995). 

For each category within each variable there are three possible values based all 

of them in the distribution of the variables to be compared among both files and 

taking into account missing values. A value will be assigned to the pair if they 

coincided, a value will be assigned if they do not coincide and zero if one of the 

two values are missing. 

5. Restriction of comparison pairs (blocking) 

Once weights had been computed, it is necessary to compare the information 

obtained for the variables common to both files. This first step, known as the 




blocking phase (Jaro, 1995) consisted in forming blocks in order to reduce the 

comparisons number. In our case we form blocks with those HDR records for 

which the date and time of patient attendance was within three days after the 

crash occurrence reported in the police files. 

6. Comparison stage (matching) 

Within each block, two level comparisons were made: firstly, the contents of the 

common variables for both files for each HDR record with each DGT record 

were compared. 

7. Simple weights assignment 

Out of every between-variable comparison a weight value was assigned. 

8. Computation of composite weights 

At a second stage, a composite weight as the sum of the individual weights 

obtained in between-variable comparisons was generated, allowing the 

comparison between records. 

Figure 42 indicates that the distribution 

of the variable WEIGHT is bimodal. It 

seems reasonable to think that some 

characteristic in the registries exists that 

divides them in two groups, each one 

with a normal distribution. 

We would suppose that those records 

that have not agreed with their real pair 

will have a relatively low weight, and in 

some few cases - based on the little 

frequency - not yet being real, the 

weight can get to be relatively high. On 

the other hand, those that have agreed 

with their pair, in most of cases will have 

Figure 42: Distribution of weights 

0 

0 5 10 15 20 25 

Weight 

assigned a relatively high weight, except those with the most common 

characteristics, that they will have the smaller weight. Under these assumptions, 

we can think that the characteristic that divides the records in two groups is 

being or not a real pair. The observation of the figure 7.7.2 indicates that a good 

cut off to differentiate the two distributions could be the corresponding one to 

weight 13,5 (W L ). 

80 

Fr 

ec 

ue 

nc 

y 

60 

40 

20 




9. Decision stage (linking) 

An HDR record was matched to a police record when it was the record with the 

highest composite weight after its comparison with all the remaining records in 

the selected block. If there are two or more records with the same weight for 

one HDR record, it remains unlinked, because it is impossible to distinguish 

which record corresponds to the same person. 

Out of 1693 HDR records, 1016 have been linked (63%), although we need to 

decide which of them are really true pairs. In order to decide which are true 

pairs, we analyse both the number of variables that agree and the weight of 

each pair. 

Table 107 shows the distribution of the number of coincidences. Two of the 

1,016 records obtain the maximum weight with a record with which they only 

have in common or the age, or the sex, or the day of the accident or the 

province where the accident happened. That is its “better pair”, but it is not 

sufficient to trust that it is its true pair. On the other hand, 45.6% of the cases 

were associated to a record which agreed in all the variables. 

Table 107: Distribution of number of coincidences. 

N % 

1 2 0,2 

2 112 11,0 

3 439 43,2 

4 463 45,6 

Total 1.016 100,0 

Figure 7.7.3 shows the distribution of the weights assigned according to the 

number of coincident variables. The weighs cut-off W L is also indicated. Using 

this point, all the records in which all the variables have agreed and 7% are 

considered correct of which the three variables have agreed. They do not 

consider any of the pairs with less correct than three coincident variables. 

Figure 43: Weight distribution according the number of coincident 

variables 

for n_coinc= 2 



80 

80 

80 

60 

Fre 

que 

ncy 

40 

60 

40 

60 

40 

20 

20 

20 

0 

0 5 10 15 20 25 

weight 

Weight 

0 

0 

0 5 10 15 20 25 0 5 10 15 20 25 

weight 

pes 

Weight 

Weight 




10. Threshold determination 

Using these weights, two threshold values were defined: the lower-threshold 

limit, under which all records with such weight value would be considered not 

corresponding to the same individual, and the upper-threshold limit, above 

which a record would be considered to belong to the same individual. 

11. Review of dubious pairs 

For those pairs with a weight value between the two threshold limits, a manual 

review of the data by three reviewers was established, using additional 

information, in order to decide if the linkage was accepted. The software W- 

conecta includes and adaptation of the sequential review process used in the 

field of quality control. That is when dubious cases are reviewed, considering 

the number of correct or incorrect pairs, the software indicates how the 

thresholds must be modified. If the chosen interval there are too many correct 

records the program suggest to lower the upper threshold, and if there are too 

many incorrect records to higher the lower threshold. 

The intervals decided to be reviewed were established observing the histogram 

and the distribution of weights. Review of dubious cases was done by three 

people to assure objectivity. A conservative criteria was chosen: the linkages 

between records considered correct by two or more persons, were considered 

correct linkages. The same criteria was used for incorrect pairs of records. 

We will consider as linked 493 pairs. This represents 48.5% of the HDR records 

that had assigned to some police record, and 30% of the total of HDR records 

Castilla y Leon. 

Finally, it is necessary to discard the possibility that the connected data are 

biased towards the cases of more peculiar characteristics. Table 7.7.8. shows 

the basic characteristics of the individuals in the three groups: total hospitalised, 

hospitalised which have assigned any police record (63%) and total of pairs that 

are considered truly linked (30%). 




Table 108: Main characteristics of the individuals hospitalised due to road 

injuries. Percentage, quartiles and values minimum and maximum. 

Total hospital Linked records Real pairs 

records 

N 1.636 1.016 493 

GENDER (%) 

Male 70,5 70,7 70,8 

Female 29,5 29,3 29,2 

AGE 

Mín 0 0 2 

P25 21,0 23,0 24,00 

P50 35,0 35,5 35,00 

P75 55,0 56,0 55,50 

Máx 97 97 97 

PROVINCE (%) 

Avila 4,6 5,4 7,1 

Burgos 22,7 21,2 20,9 

León 19,8 18,8 18,3 

Palencia 8,6 9,4 10,5 

Salamanca 11,6 12,2 14,2 

Segovia 4,4 4,7 3,7 

Soria 3,9 5,3 5,9 

Valladolid 16,8 15,2 12,4 

Zamora 7,6 7,9 7,1 

LENGTH OF STAY 

Mín 0 0 0 

P25 2,0 2,0 2,0 

P50 5,0 5,0 6,0 

P75 11,0 11,0 12,0 

Máx 109,0 89 82 

TYPE OF DISCHARGE (%) 

Home 85,7 83,2 81,1 

Transfer 10,5 11,6 12,8 

Voluntary discharge ,8 1,0 1,0 

Fatality 3,0 4,1 5,1 

The results obtained in the previous analyses suggest to consider as weight 

threshold (W L ) the value of 13,5. The pairs with a weight equal or greater than 

this are considered true linked pairs (corresponding to the same individual). 

Validity of record linkage 

As there are no unique identifiers or a subsample of records with unique 

identifiers that could identify the linked pairs as real pairs, two fictitious data 

bases for the calculation of the values of Sensitivity and Specificity have been 

built up. 

The databases were created using the following procedures: initially we decide 

the percentage of cases of the source that we expect to link with the records of 




source B. In our case, we suppose that we would expect to be able to link 70% 

the records of the HDR with any record of the DGT. 

The fictitious database contains the HDR records used for the linkage (n=1636). 

Out of these records 70% have been selected randomly that will also be 

included in the data base B. There is a variable that identifies each record (ID) 

and another that indicates if the registry comes also from data base B. 

The fictitious database B contains 70% of the records used for the linkage that 

we have selected randomly (n=1.187) and in addition, have added 5,800 

records corresponding to other years which the year of the accident has been 

modified until arriving to quadruplicate the size of the data base A (from the data 

of previous years, we know that the record number of the data base of the HDR 

is approximately one fourth of the record number of the DGT). A new ID has 

been assigned to the new cases different from those assigned previously to the 

records of source A. 

Once the two databases have been created, we proceed to link them. Records 

are classified as linked and unlinked according to the value of the weight 

threshold (W L ) established in the real connection. The ID variable is used to 

judge the real status of the pair of registries. Thus, we will have four possible 

situations (Table 109): 

Table 109: True and false positives and negatives according the 

coincidence with the identifier (ID) 

Real pair (same ID) 

No real pair (different ID) 

Result of 

linkage 

process 

linked 

True Positives 

W ≥ W L 

(TP) 1,129 

Not linked False Negatives 

W< W L (FN) 58 

False Positives 

(FP) 20 

True Negatives 

(TN) 429 

True Positives (TP): Number of records pairs linked correctly. 

False Positives (FP): Number of records pairs wrongly linked. 

True Negatives (TN): Number of records unlinked pairs correctly. 

False Negatives (FN): Number of records wrongly unlinked pairs. 

From the previous parameters, we can obtain different measures to evaluate 

the accuracy the record linkage: 

• Sensitivity (S): S=TP/(TP+FN) 

The number of pairs of records true linked pairs divided by the total number of 

correct pairs of records. It is interpreted as the probability that a concordant pair 

of records has been connected by the process. 




• Specificity (E): E=TN/(FP+TN), 

The number of unlinked correctly divided the total number of pairs of incorrect 

records. It is interpreted as the probability that a discordant pair does not 

connect in the process. 

• Positive Predictive Value (PPV). VPP=TP/(TP+FP) 

The number of correct linked records divided by the total number of pairs of 

linked records. It is interpreted as the probability that a pair linked is a really 

pair. The PPV is useful as an indicator accuracy of the linkage process. 

• Match Rate (MR). MR=(TP+FP)/(TP+FN) 

The total number of linked records pairs divided by the total number of pairs of 

correct records. 

• From our data, out of 1,636 HDR records, 1,149 were linked with a 

weight equal or higher than 13.5. It yields a positive predictive value of 

98.3%. 

• Out of 1,187 records of both databases, 1,129 were linked, which yields 

a sensibility of 95,1%. 

• Out of 449 records of A database 429 were not linked, which yields a 

specificity of 95,5%. 

• The match rate was 0,97. 

According to this test there would be 20 false positives and 58 false negatives. 

This corresponds to 0.3% and 0.8% respectively of the police database, and 

1.2% and 3.5% of the HDR database. We conclude that the likelihood that a 

pair of linked records belongs to the same person and accuracy of the 

procedures are high. 

7.7.4 Results 

In Castilla y Leon, during July-December of 2005, Police reported 6,970 people 

injured in traffic accidents, and public hospitals reported 1,636 persons 

hospitalised due to road injuries. Four hundred and ninety three records were 

linked between both databases. This corresponds to 7,1% of police records and 

30% of hospital records. As expected the proportion of serious casualties is 

higher among linked records (Table 7.7.11.) 




Table 110: Linked and unlinked police and hospital records distribution. 

Police 

Hospital 

n % n % 

Yes 493 7.1 493 30.1 

No 6477 92.9 1143 69.9 

Total 6970 100.0 1636 100.0 

Table 111: Linked and unlinked police records by severity. 

Police Linked Unlinked Total 

severity n Col % Row % n Col % Row % n Col % Row % 

Fatal 13 2,6 5,3 234 3,6 94,7 247 3,5 100.0 

Serious 326 66,1 20,3 1283 19,8 79,7 1609 23,1 100.0 

Slight 154 31,2 3,0 4933 76,2 97,0 5087 73,0 100.0 

Unknown 27 0,4 100,0 27 0,4 100.0 

Total 493 100 6477 100.0 6970 100.0 100.0 

The distribution of casualties by MAIS is similar among linked and unlinked 

hospital records. The proportion of linked records increases with severity, from 

30.9% for MAIS 1 to 50% for MAIS 6. 

Table 112: Linked and unlinked hospital records by MAIS (Maximum 

Abbreviated Injury Severity). 

MAIS Linked Unlinked Total 

n Col % Row % n Col % Row % n Col % Row % 

0 20 4.1 17,4 95 8.3 82,6 115 7.0 100.0 

1 54 11.0 30,9 121 10.6 69,1 175 10.7 100.0 

2 242 49.1 29,7 572 50.0 70,3 814 49.8 100.0 

3 99 20.1 31,8 212 18.5 68,2 311 19.0 100.0 

4 51 10.3 33,3 102 8.9 66,7 153 9.4 100.0 

5 20 4.1 40,0 30 2.6 60,0 50 3.1 100.0 

6 1 0.2 50,0 1 0.1 50,0 2 0.1 100.0 

9 6 1.2 37,5 10 0.9 62,5 16 1.0 100.0 

Total 493 100.0 30,1 1143 100.0 69,9 1636 100.0 100.0 

The proportion of linked records increased also with longer length of stay. 

Table 113: Linked and unlinked hospital records by length of stay. 

MAIS Linked Unlinked Total 

n Col % Row % n Col % Row % n Col % Row % 

Overnight 15 3.0 24.6 46 4.0 75.4 61 3.7 100.0 

1-3 149 30.2 25.9 426 37.3 74.1 575 35.1 100.0 

>3 329 66.7 32.9 671 58.7 67.1 1000 61.1 100.0 

Total 493 100.0 30.1 1143 100.0 69.9 1636 100.0 100.0 




The distribution of gender among linked records is similar to hospital only 

records, but the proportion of females is 5% lower than police only records. 

Regarding age, linked records shows a higher proportion (7.3%) of children 

under 14 years old than police only records (4.5%), but lower than hospital only 

records (14%). Youth from 18 to 35 years are more represented among police 

only records (44.8%), and less among hospital only (29.6%) and linked records 

(39.6%). On the other side, elderly are more represented among hospital only 

records (15.1%) and linked records (15,6%). 

Table 114: Gender and age group according to data source 

Police only Police ∩ Hospital Hospital only 

N=6477 % N=493 % N=1143 % 

Gender 

Male 4141 63.9 344 69.8 805 70.4 

Female 2274 35.1 148 30.0 338 29.6 

Unknown 62 1 0.2 

Age 

0-13 years 290 4,5 36 7,3 160 14,0 

14-15 years 109 1,7 8 1,6 41 3,6 

16-17 years 199 3,1 10 2,0 40 3,5 

18-35 years 2900 44,8 195 39,6 338 29,6 

36-50 years 1433 22,1 94 19,1 221 19,3 

51-65 years 814 12,6 73 14,8 170 14,9 

66-98 years 602 9,3 77 15,6 173 15,1 

999 130 2,0 

Information about the type of road user or of vehicle is not available for 

hospitals. It should be recorded with the E-code of the ICD-9-CM. But usually it 

is not recorded, or it is recorded only as traffic crash, without specifying anything 

else. A previous study (Pérez et al, 2006) showed that in Spain the E-code only 

gives useful information on these variables for 21% of HDR at national level. 

Therefore it is not possible to compare and to derive conversion factors for road 

user or vehicle. 

Among the linked records there is a higher proportion of pedestrians, pedal 

cyclists and motor cyclist , and a lower proportion of car occupant (Table 115 

and Table 116). 




Table 115: Road user distribution among linked and unlinked records 

Police only Police Linked 

n % n % 

Car occupant 4389 67,8 286 58,0 

Pedestrian 440 6,8 64 13,0 

Pedal cyclist 112 1,7 13 2,6 

Motor cyclist 636 9,8 68 13,8 

Other 894 13,8 62 12,6 

Unknown 6 0,1 

Table 116: Type of vehicle distribution among linked and unlinked records 

Police only Police Linked 

Type of vehicle n % n % 

Car 4389 67,8 286 58,0 

Motorcycle 290 4,5 37 7,5 

Moped 346 5,3 31 6,3 

Bicycle 112 1,7 13 2,6 

Bus 19 0,3 4 0,8 

Truck or van 706 10,9 45 9,1 

Other 169 2,6 13 2,6 

Unknown 446 6.9 64 13.0 

Some of the differences found between linked and unlinked data might be due 

to the record linkage process. Cases with less common characteristics receive a 

higher weight than those that are more common, such 18 to 35 year old men or 

car occupants. It is more likely to find several records with the same 

characteristics. 


Table 117 and Table 119 show the number of cases reported by police and 

hospital and the estimated cases according police severity by length of stay and 

MAIS. Table 118 and Table 120 show the conversion factors derived from these 

data. (Results are also presented in the general section of the report). 





Stay 

Table 117: Police and hospital reported cases and estimated cases by 

length of stay 


Police severity 

Not 

known 

Not in 

police 

database 

Total Fatal 

Estimated cases 

Serious Slight Total 

Overnight 7 3 5 46 61 28 12 20 61 

1-3 nights 3 80 66 426 575 12 309 255 575 

>3 nights 3 243 83 671 1000 9 739 252 1000 

Sub total 13 326 154 1143 1636 43 1082 511 1636 

Police not 

hospital 

234 1283 4933 27 234 1283 4933 6450 

Total 247 1609 5087 27 1143 8113 277 2365 5444 8086 

MAIS 

Table 118: Conversion factors by length of stay 




Overnight 0.115 0.008 0.004 

1-3 nights 0.047 0.192 0.050 

>3 nights 0.037 0.459 0.050 

Sub total 0.175 0.672 0.100 

Police not hospital 

Total 1.122 1.470 1.070 

Table 119: Police and hospital reported cases and estimated cases by 

MAIS 

Fatal 


Police severity 

Not 

known 

Not in 

police 

database 

Estimated cases 

Total Fatal Serious Slight Total 

1+2 4 178 114 693 989 2471 18781 5314 1 7439 

3 0 81 18 212 311 0 254 57 311 

4 3 40 8 102 153 9 120 24 153 

5 4 14 2 30 50 10 35 5 50 

6 0 0 1 1 2 0 0 2 2 

0+9 2 13 11 105 131 10 66 55 131 

Sub total 13 326 154 1143 1636 276 2353 5457 8086 

Police not 

234 1283 4933 27 

hospital 

Total 247 1609 5087 27 1143 8113 

1 

Includes proportional distribution by severity plus “Police not hospital” cases, 

assuming that these cases are of low severity. 




Table 120: Conversion factors by MAIS 

MAIS 



1+2 1,001 1,167 1,045 

3 0,000 0,158 0,011 

4 0,036 0,075 0,005 

5 0,040 0,022 0,001 

6 0,000 0,000 0,000 

0+9 0,041 0,041 0,011 

Total 1,119 1,462 1,073 


In this report we have shown that although it has been impossible to carry out a 

national study to estimate conversion factors to address the issue of 

underreporting road casualties, it is feasible to carry out the study with only one 

autonomic region. The validity of the record linkage procedure is quite good, 

and the conversion factors derived seem to yield reasonable estimates at 

national level. 

The proportion of police records linked has been very low (7,1%). This 

proportion was expected because we are linking with hospitalised records, 

which by definition have some level of severity, and not all casualties. We know 

from a Road Injuries Information System based on hospital emergencies, that in 

Barcelona 7,8% of road casualties are hospitalised and 3,2% are transferred to 

another hospital, some of them are later on hospitalised. 

Conversion factors derived from this study would not be appropriate for slight 

and for urban casualties, as slight casualties are underreported in the Castilla y 

Leon data. We tried to estimate conversion factors from a record linkage done 

in Barcelona. But it is not representative of the national urban data, because in 

Barcelona local police report exhaustively traffic crashes and therefore 

conversion factors were very small. This is not the situation in many cities in 

Spain, although the quality of reporting is improving because many cities are 

setting up urban road safety plans. 

Some limitations need to be considered. First of all, due to lack of availability 

data we only used a 6 months database from only one autonomous region. It 

would be convenient to repeat the study including a year or even several years 

of data and at least one or two more autonomous regions where the study 

would be feasible and were there is a good balance of urban and non urban 

crashes. 

Secondly, we do not have information about the coverage of HDR in Castilla y 

Leon. For the whole Spain, we know that for the recent years the HDR has a 

good coverage of public hospitals, around 99%, but do not include private 

hospitals. We can assume, however that serious injuries in general attend a 

public hospital. Nonetheless it should be assessed. 




Thirdly, and related to coverage when studying a region, it can happen that in 

some cases the persons injured in a collision in Castilla y Leon might be 

transferred to a neighbourhood autonomous region because there is a large 

hospital closer. This can occur in for instance in provinces closer to Madrid. In 

this case police records would not be able to be linked to hospital databases, 

and conversion factors would give lower estimates. 

Finally it would be very useful to derive conversion factors from different 

features such gender an age, and type of road user. The E-code is recorded 

better In some autonomous regions. 

In conclusion, the methodology of the study carried out to obtain conversion 

factors to estimate serious casualties is appropriate, but in order to apply these 

conversion factors we recommend to repeat the study with a broader database. 

Acknowledgements 

To Eva Cirera and Marc Marí-Dell’Olmo. To the Department of Information and 

Health from Castilla y León and the National Traffic Safety Authority (DGT). 


Arribas P, Cirera E, Tristan-Polo M. (2004). Buscando una aguja en un pajar: 

las técnicas de conexión de registros en los sistemas de información sanitaria. 

Med Clin (Barc). 2004 Feb 15;122 Suppl 1:16-20 

Cirera E, Plasencia A, Ferrando J and Arribas P (2000). Probabilistic linkage of 

police and emergency department sources of information on motor-vehicle 

injury cases: a proposal for improvement. J Crash Prevention and Injury Control 

2000; 2(3):229-237. 

Cook LJ, Olson LM, Dean JM. Probabilistic record linkage: Relationships 

between file sizes, identifiers, and match weights, 2001;40:196-203. 

Roos LL, Wajda A. Record linkage strategies. Methods of information in 

Medicine, 1991;30:117-123 

Jaro M A. Matching and Record Linkage, in Business Survey Methods, eds. B. 

G. Cox, D. A. Binder, B. N. Chinnappa, A. Christianson, M. J. Colledge, and P. 

S. Kott, New York: Wiley, 1995 pp. 355–384. 

MacKenzie, E. J., Sacco, W et al. (1997). ICDMAP-90: A users guide. 

Baltimore, The Johns Hopkins University School of Public Health and Tri- 

Analytics, Inc. 

Newcombe HB (1988). Handbook of record linkage: methods for health and 

statistical studies. Administration and business. Oxford: Oxford University 

Press,. 




Pérez C, Cirera E, Borrell C, Plasencia A on behalf of the work group of the 

Spanish Society of Epidemiology on the Measuring of the Impact on health of 

road traffic accidents in Spain. Motor vehicle crash fatalities at 30 days in Spain. 

Gac Sanit. 2006; 20(2):108-15. 




7.8 Study carried out in the United Kingdom 

Report prepared by Jeremy Broughton and Maureen Keigan (TRL) 


The UK study has consisted of linking the road accident data from Scotland for 

1997-2005 with medical data from the Scottish Hospital In-Patient System. The 

linkage had been carried out previously for road accidents occurring between 

1980 and 1995 (Stone, 1985; Keigan et al, 1999). For the SafetyNet project, the 

procedures and software were updated and applied to data from the 1997-2005 

period. Various comparisons are made of the results from the current study with 

the results of these earlier studies. 

The general concept of linking and comparing road accident and medical 

databases has been applied in a number of previous studies and can be 

implemented in various ways. The study reported by Simpson (1996), for 

example, used medical data recorded by clerks working in A&E Departments of 

a sample of 16 hospitals in Great Britain. Collecting data in this way is relatively 

expensive, however, so the specific approach adopted for the SafetyNet 

collaboration was selected on the basis of the funding available and the level of 

experience that existed in the eight national teams. 

The United Kingdom comprises England, Wales, Scotland and Northern Ireland. 

Scotland is the most northerly of these countries, with almost 9% of the overall 

population and total of registered vehicles. While Scotland has had a devolved 

government for several years, the same traffic laws apply throughout the United 

Kingdom, the 8 Scottish police forces operate in the same way as those in the 

rest of the country and they are subject to the same operational pressures. 

Table 121 compares the casualties in Scotland in 2005 with the UK total. 

Table 121: Proportion of UK casualties in Scotland, 2005 

UK total % in Scotland 

killed injured killed injured 

Pedestrians 699 33249 9.4% 9.0% 

Pedal cyclists 152 16558 10.5% 4.6% 

Motorcycle users 584 24669 5.8% 4.2% 

Car users 1756 182797 8.8% 6.1% 

Others 145 18567 11.0% 8.6% 

All road users 3336 275840 8.6% 6.3% 


The STATS19 file 

All road accidents involving personal injury and at least one vehicle occurring on 

the highway ('road' in Scotland) that were reported to the police within 30 days 

are recorded in the National Road Accident database (STATS19). Details of 

accident circumstances and the vehicles and casualties involved are recorded 

in annual files. 




A file of all casualties in Scotland was extracted from the national STATS19 files 

for the years 1997 to 2005. SHIPS data were supplied for 1995-2005, but the 

ICD9 system of injury coding used in 1995 and 1996 was superseded in 1997 

by the ICD10 system so the main results are from 1997-2005. 

The STATS19 variables extracted for the matching process consisted of police 

force area, date and time of the accident, casualty type, age and gender, 

casualty severity, road user and type of vehicle involved. The type of road user 

and vehicle involved were combined to create the following road user classes 

which were used for matching. 

Code 

Road user 

1 Driver of a motor vehicle 

2 Passenger of a motor vehicle 

3 Rider of a motorcycle 

4 Passenger of a motorcycle 

5 Pedal cyclist 

6 Pedestrian 

7 Unknown 

The SHIPS file 

The Scottish Hospital In-Patient System (SHIPS) data were supplied by the 

Healthcare Information group of NHS National Services Scotland. It was 

released to TRL under a confidentiality statement for users of NHS patient data. 

Episodes (casualties) with an 'Emergency - Road Traffic Accident' type of 

admission or a specified Road Traffic Accident diagnostic code on the hospital 

discharge that were admitted and discharged within the years 1997 to 2005 

were selected. 

TRL has been advised that the operational procedures for SHIPS were 

unchanged between 1997 and 2005, and indeed for many years previously. 

Thus, any changes that are identified when the data are analysed by year 

cannot be attributed to changes in the data collection procedure. They must 

caused by changes in the number and nature of casualties, or the criteria used 

for admission as a hospital in-patient. 

A large number of variables are provided within this file including hospital code, 

age and gender of the patient, admission type, length of stay and six diagnosis 

codes. The hospital code is a unique alpha-numeric five character code. The 

age of the in-patient is provided and has been calculated from date of birth. The 

admission type notes the reason for admission and includes, for example, 

whether it is from an emergency, a transfer or the waiting list. The length of stay 

used in this report refers to total length of stay and is either for one admission or 

has been accumulated from a number of 'stays' in connection with the 

diagnoses. Thus, the earlier problems (Stone, 1984) of linking patient records 

for the same person with more than one admission or with additional transfer 

records from the same accident have been eradicated by the suppliers of the 

data. 

The files sent to TRL now hold information on a Continuous Inpatient Stay 

basis. A continuous inpatient stay (CIS) is a continuous period of time spent as 




an inpatient or day case in hospital regardless of any transfers between 

specialities or hospitals. From 1997 the diagnosis codes are based on the 

International Classification of Diseases (ICD) codes (World Health Organisation, 

1992) using the ICD10 format. 

Length of Stay in the SHIPS data is in effect the number of nights spent in 

hospital. 0 days is recorded for a patient admitted and discharged on the same 

day, 1 day for a patient discharged the day after admission (irrespective of time 

on either day) etc. It is thus possible to adopt exactly the definition of Length of 

Stay set out in Section 2.2. 

Some of the variables included in the SHIPS data have been used to generate 

fields for use in matching the two datasets. The ICD10 V codes relate to 

external causes of morbidity and mortality; codes V01 - V99 define transport 

accidents and have been used to determine the mode of transport of the 

casualty. These codes, provided in the SHIPS data as one of the six diagnosis 

codes, are as follows. 

ICD10 V code 


V01 - V09 

Pedestrian 

V10 - V19 

Pedal cyclist 

V20 - V29 

Motorcyclist 

V30 - V39, V80 - V86 Other motor vehicle 

V40 - V49 

Car occupant 

V50 - V59 

Light goods vehicle 

V60 - V69 

Heavy goods vehicle 

V70 - V79 Public service vehicle occupant 

V87 

Other vehicle 

V89, V98 - V99 Vehicle type unknown 

The V codes also have a fourth character subdivision which has been used, 

where possible, to assign the casualty as a driver or passenger of the vehicle. In 

those records where a V code was not present a road user class of unknown 

has been assigned. 

A look-up table was developed to achieve common definitions of road user. The 

seven classes of road user defined using STATS19 variables were taken and 

assigned to the V codes in ICD10. This set of codes has been used previously 

for linking data from these two sources for years prior to 1996. The rationale for 

using them again was that this would enable comparisons to be made of 

proportions of records that have been linked for datasets from previous years, 

thus ensuring confidence in this revised matching method. 

The hospital code present in the SHIPS data provided a link to the police force 

area held within STATS19 as follows. As would be expected, the larger police 

forces have more hospitals within their areas. 




Police Force Area 

Northern 

Grampian 

Tayside 

Fife 

Lothian and Borders 

Central 

Strathclyde 

Dumfries and Galloway 

Hospital code letter 

W, R, H, Z 

N 

T 

F 

B, S 

V 

C, G, A, L 

Y 


The principles used in this method have been developed from the work carried 

out by Nicholl (1980) and Stone (1984). In the first instance a 10 per cent 

sample of Hospital In-Patient Enquiry (HIPE) records were matched with 

STATS19 cases for England and Wales. This produced a matching procedure 

with a success rate of about 50% using distance from a hospital with an 

accident and emergency department, date and time (for accidents that occurred 

on the day before admission), gender, age to produce a score for each match. 

This work recommended that the E-code from the ICD9 code, which represents 

class of road user, should be made available for further work to reduce the 

number of possible matches. 

The study reported by Stone (1984) is a comprehensive matching of police 

accident records and 100% sample of hospital in-patients records for Scotland 

for the year 1980. A matching algorithm using a similar basis to Nicholl in terms 

of matching variables and degrees of tolerance was developed. The bonus for 

this study was that the E-code describing class of road user was available and a 

matching proportion of 67% was achieved. Scottish Hospital In-patient data has 

been matched to STATS19 using this method for the years 1980 to 1995. 

The main variables used in the matching process for the data reported here 

were police force area/hospital code, date of accident/admission, casualty/inpatient 

age and road user class. In addition, some STATS19 variables were 

allowed to vary according to the tolerance level; these are casualty severity, 

date of accident, police force area, age and road user class. Gender was 

included as a matching variable and was always an exact match. 

Method 

The data were imported into a MS Access database and queries have been 

conducted to perform the matching process. The tolerance levels have been 

applied in ascending order and once a record has been matched then it is 

withdrawn from further matching. The lower tolerance levels are considered the 

best match, for example, a match at tolerance level 1 requires all of the 

matching fields to have the same value. 




Tolerance values 

Using an Access update query for each of the 30 tolerance levels, the matching 

fields from the table containing the STATS19 records are linked with the 

corresponding field in the SHIPS data table. For some tolerance levels the 

STATS19 matching variables are permitted to vary, for example, in tolerance 

levels 7, 8 and 9 the age of the casualty in the STATS19 record may be plus or 

minus 1. Also, where the casualty severity is given a value of -1 a match with a 

fatal casualty in STATS19 is permitted. These queries are run to update the 

SHIPS record with the STATS19 unique identity reference and the tolerance 

level field. The tolerance values are summarised in Table 122; these are 

numeric ranges which are discussed following the Table. The procedures and 

values were established in a rigorous series of tests (Stone, 1984) and have 

been used in several STATS19-based linkage studies (Stone, 1984, Keigan et 

al., 1999, Broughton et al, 2001). 

Table 122: Summary of tolerance levels 

Level Police Age Road user Casualty Length of Hour Day 

Force area class severity stay 

1 0 0 0 0 0 0 0 

2 0 0 7 0 0 0 0 

3 0 0 0 -1 0 0 0 

4 0 0 7 -1 0 0 0 

5 0 0 0 0 0 2 1 

6 0 0 7 0 0 2 1 

7 0 1 0 0 0 0 0 

8 0 1 7 0 0 0 0 

9 0 1 0 -1 0 0 0 

10 0 1 7 -1 0 0 0 

11 0 0 0 0 0 4 1 

12 0 0 7 0 0 4 1 

13 9 0 0 0 0 0 0 

14 9 0 7 0 0 0 0 

15 0 0 0 1 5 0 0 

16 0 0 7 1 5 0 0 

17 0 2 0 0 0 0 0 

18 0 2 7 0 0 0 0 

19 0 0 5 0 0 0 0 

20 0 0 0 0 0 4 1 

21 0 0 7 0 0 4 1 

22 0 0 0 1 0 0 0 

23 0 0 7 1 0 0 0 

24 0 0 0 0 0 0 1 

25 0 0 7 0 0 0 1 

26 0 0 0 1 0 1 1 

27 0 1 0 1 0 0 0 

28 0 1 7 1 0 0 0 

29 0 3 0 0 0 0 0 

30 0 3 7 0 0 0 0 

There are certain special codes in Table 122: 




o For Police Force area a value of 9 allows a match with defined 

neighbouring areas, 0 is the exact match value. 

o Road user class unknown (value 7) allows an unknown road user class 

in SHIPS to match with any road user class in STATS19, also at 

tolerance level 19 the value 5 for road user class allows a pedal cyclist 

or pedestrian in SHIPS to match with either a pedal cyclist or pedestrian 

in STATS19 data. The exact match value is 0. 

o Casualty severity of -1 permits a match with a fatal casualty in 

STATS19. A value of 1 allows a match with a slight casualty in the 

STATS19 data. The exact match value is 0. 

o Where the length of stay is quoted as 5 a SHIPS casualty with a stay of 

5 days or less may be matched with a slight casualty. 

o The date of the accident is recorded in both datasets. However, time is 

only recorded in the STATS19 data so the hour is permitted to change in 

STATS19 by plus or minus n (1, 2 or 4) hours. The exact match value is 

0. 

o When the time is within n hours of midnight then the day is allowed to be 

the previous day and if n (1, 2 or 4) hours after midnight then the 

following day may be matched. The exact match value is 0. 

7.8.4 Results 

The SHIPS dataset includes 47,297 records for the years 1997 to 2005; these 

were matched to the STATS19 data using the 30 tolerance levels, and a total of 

26,625 (56%) matches were achieved. All of these matches have been 

allocated on a one-to-one basis using the first appropriate match. In previous 

matching studies, a number of multiple matches were obtained that needed 

manual sifting to identify the best match using the local authority area. This 

system of tolerance levels inevitably leads to matches at the higher levels being 

less precise; however, it maximises the number of matches in cases where 

several potential matches have similar details. 

The proportions of matches achieved at the different tolerance levels in 1997- 

2005 are given in Table 123. Results from the previous matching for 1993 and 

1995 are included to examine the consistency of the new linkage with the 

original linkage procedure. 




Table 123: Proportion of SHIPS records matched to STATS19 

ICD9 ICD10 

1995 

1993 Old New 1997 1998 1999 2000 2001 2002 2003 2004 2005 

method method 

A 3615 3610 3972 3469 3457 3191 2995 3016 2879 2595 2535 2488 

B 5399 5321 6096 5767 5751 5448 5341 5359 5178 4914 4837 4702 

A/B 67.0% 67.8% 65.2% 60.2% 60.1% 58.6% 56.1% 56.3% 55.6% 52.8% 52.4% 52.9% 

A=number of matched SHIPS records, B=number of SHIPS records 

In addition, new data for 1995 was obtained (the original 1995 dataset was no 

longer available at TRL) and the new matching procedure was applied; the data 

allow the results of the previous and the new method of matching to be 

compared directly. The number of SHIPS records for 1995 in this dataset was 

6096 compared to 5321 previously obtained, the matches at the individual 

tolerance levels are broadly similar apart from level 24 where the proportion 

doubles from 2 to 4.1%, this tolerance level allows a SHIPS casualty to match 

with a STATS19 record on the following day (all other variables are exact). The 

overall proportion matched using the new procedure was slightly less at 65.2% 

compared to 67.8%, which is indicative that the method of linking has been 

applied consistently. 

However, there is a steady fall in the proportion of SHIPS records that have 

been matched to STATS19 records during the years 1997 to 2005. This could 

indicate an overall decline in the proportion of road accident casualties reported 

to and by the police. Alternatively, certain attributes of an accident may make it 

more or less likely to be reported, and the range of casualties may have 

changed over this period so as to reduce the proportion that is likely to be 

reported to and by the police. For example, an increasing proportion of 

casualties could be admitted to hospital for observation, but found to be 

uninjured. 

The proportion of matches achieved when the road user class is unknown in 

SHIPS and allowed to match with any class of road user in STATS19 is lower 

for the period 1997 to 2005 than for the earlier years. This may be a 

consequence of the change from using ICD9 to ICD10. 

The ICD10 codes appear to determine the class of road user more precisely 

than the ICD9 codes. The proportion of SHIPS casualties where the road user 

class was unknown from the ICD9 code was almost 3 times higher than for the 

years where ICD10 codes were used. 

The earlier version of the matching procedure ended with a manual comparison 

of the unmatched records, which identified further potential matches that lay just 

outside the system of tolerance levels. This was a laborious and time-consuming 

process that involved a degree of subjectivity. This final step was not 

included with the new matching procedure, which partly explains the slightly 

lower matching levels achieved by the new procedure – especially for 1995. 

It is planned to systematically assess a subset of unmatched records to see 

whether the system of tolerance levels can be improved. In particular, there are 




indications that police reporting of age may have tended to deteriorate over 

time, so it may be appropriate to relax the age-related tolerances. 

The AIS scores for each casualty are estimated from the ICD codes in the 

SHIPS file. The SHIPS data from 1997 used the ICD10 system, whereas the 

ICD9 system had been used previously, so a new mapping from ICD to AIS was 

needed. The mapping from ICD10 to AIS98 developed at the University of 

Navarra (Apollo, 2006) was adopted. 

Whether MAIS is estimated directly or indirectly via such a mapping, 

unexpected results can often be found, such as fatal casualties with low MAIS 

scores. This may appear to call into question the validity of the linkage, although 

such results are also found in studies which do not rely on record linkage. 

Consequently, a number of such cases were examined in detail, and in each 

case the outcome was consistent with the ICD codes, i.e. there was no cause to 

doubt the linkage. Four SHIPS cases with MAIS 2 that were linked to a 

STATS19 fatality are presented below: 

Trends 

Male pedestrian aged 20, MAIS 2, ICD10 injury codes included: 

Intracranial injury, unspecified injury of abdomen, lower back and pelvis 

and cardiac arrest. Length of stay was 0 days. 

Male pedestrian aged 83, MAIS 2, ICD10 injury codes included: fracture of 

vault of skull, unspecified injury of thorax, atrial fibrillation and flutter and 

respiratory failure (unspecified). Length of stay was 3 days. 

Male driver aged 19, MAIS 2, ICD10 injury codes included: other 

unspecified injuries of multiple body regions and cardiac arrest. Length of 

stay was 1 day. 

Male driver aged 34, MAIS 2, ICD10 injury codes included: intracranial 

injury and dependence on respirator. Length of stay was 19 days. 

Keigan et al. (1999) presents various tables of results for 1980-95 that allow 

comparisons to be made with results of the new linkage, and overall trends 

since 1980 will now be examined. Some comparisons can be made exactly, 

others must be approximate because of the groups used in the earlier report. 

First, Figure 44 examines the proportion of STATS19 casualties in Scotland that 

could be linked to SHIPS records. The data for 1980-95 come from the previous 

linkage, the data for 1997-2005 come from the new linkage and there has been 

no linkage for 1996. Between 53 and 60% of serious STATS19 casualties each 

year were linked to SHIPS records; the mean proportion in 1997-2005 is 57.5%, 

slightly higher than the 1987-95 mean of 56.8% although there has been a 

downward trend since 1999 of about 0.7% p.a. 




Figure 44: The proportion of STATS19 casualties linked to SHIPS 

70% 

60% 

50% 

40% 


30% 

20% 

10% 

0% 

1980 1985 1990 1995 2000 2005 

Figure 4 appeared in Section 4.1 of the main report. It compares the distribution 

of Length of Stay in the linked casualty data with three ranges: 0 days (admitted 

and left hospital on the same day), 1-3 days and over 3 days. This uses the 

definition of Length of Stay that was applied in the earlier study, for consistency, 

so the data from the current study have been recalculated. There are clear 

overall trends, with a shift towards shorter stays in hospital. The changes 

between 1991-95 and 1997-99 fit broadly within the overall trends, so it appears 

that the results of the new linkage are consistent with those of the original 

linkage in terms of Length of Stay. 

Figure 5 also appeared in Section 4.1 of the main report, and presents the 

corresponding comparison for MAIS. This comparison is affected by the switch 

from ICD9 to ICD10. A new mapping from ICD10 to AIS98 developed at the 

University of Navarra (Apollo, 2006) is used, and it produced an appreciable 

proportion of MAIS 9 (unknown) scores. These appear to be generally minor 

injuries, so they have been included with MAIS 1 to prepare the Figure. The 

Figure shows major increases in the proportion of casualties with MAIS 1 

between 1991-95 and 1997-99, and corresponding reductions with higher MAIS. 

If, less plausibly, the MAIS 9 scores are distributed pro rata among the known 

codes then the changes are a little less definite, but it is clear that the 

combination of the ICD10 codes and the new mapping has tended to yield lower 

MAIS scores. There is no way of telling whether the earlier or the later system 

yields the more reliable results, but this Figure does indicate that results based 

on mapping ICD9 codes to MAIS should not be compared with results based on 

mapping ICD10 codes. 

The relationship between MAIS and Length of Stay is examined in Table 124 for 

the 1997-2005 period (MAIS 9 is included with MAIS 1, serious and slight 

casualties only are included). For example, 15% of MAIS 1 casualties left 

hospital on the same day in 1997-99, compared with 17% in 2003-05. The 




distributions vary markedly by MAIS, and there is a trend for shorter hospital 

stays at most MAIS levels. 

Table 124: Distribution of casualties by Length of Stay at each MAIS level 


MAIS overnight 1-3 days >3 days 

1 1997-99 15% 74% 11% 

2000-02 17% 74% 10% 

2003-05 17% 74% 9% 

2 1997-99 4% 51% 45% 

2000-02 5% 51% 45% 

2003-05 5% 48% 47% 

3 1997-99 1% 13% 86% 

2000-02 1% 12% 87% 

2003-05 1% 15% 84% 

4-6 1997-99 1% 3% 96% 

2000-02 2% 6% 93% 

2003-05 1% 9% 90% 

This analysis is extended in Table 125 to compare the distributions for the 1997- 

2005 period by road user type; pedal cyclist and other casualties are omitted 

because of the smaller casualty numbers. The Table shows that car occupants 

tend to be discharged slightly earlier than pedestrians at each level of MAIS, 

while motorcyclists tend to spend longer in hospital. 

Table 125: Distribution of casualties by Length of Stay by road user type, 

1997-2005 


MAIS Road user type overnight 1-3 days >3 days 

1 car occupant 18% 72% 10% 

pedestrian 13% 77% 10% 

motorcyclist 9% 79% 13% 







4-6 car occupant 3% 6% 92% 



Cost-benefit analysis 

One way of understanding the relative contribution of a particular group of 

casualties to the national total is via cost-benefit analysis. For example, Table 

126 uses the British Government’s cost-benefit value of prevention of road 

accidents to show the relative contribution of fatal, serious and slight casualties 




to the national cost of injuries. It includes the number of serious* casualties 

(MAIS≥3) estimated using the conversion factors from the STATS19/SHIPS 

linkage (Table 8). The official values are averages per severity, not by MAIS, 

and it is assumed in the Table that the value of a serious* casualty is twice the 

official value of a serious casualty. This is probably conservative, as this 

category includes those who died more than 30 days after the accident and 

those who suffered long-term disability. 

Table 126: Cost-benefit value of prevention of road accidents, Great 

Britain, 2005 

Cost per 

casualty (£m) 


casualties Cost (£m) 

Fatal 1.42846 3201 4573 38% 

Serious 0.16051 28954 4647 38% 

Slight 0.01238 238862 2957 24% 

Total 12177 100% 

Serious* 0.32102 6945 2230 18% 

Cost per serious* casualty estimated as twice cost per serious casualty 

The lower threshold of the serious injury category in Great Britain is set 

relatively low, but it appears that non-fatal casualties with MAIS≥3 account for at 

least one half of the burden of serious casualties, and one fifth of the overall 

injury burden. 


The matching of SHIPS and STATS19 records that had previously been carried 

out at TRL for 1980-95 has been successfully extended to the 1997-2005 

period. This involved developing an ACCESS database that applied the 

principals of the earlier matching procedure. The checks described above 

suggest that the two procedures are broadly consistent. Table 123 showed that 

the proportion of SHIPS records that could be matched to STATS19 has fallen 

steadily since at least 1993, and it is planned to investigate this further. One 

possible explanation would be that changes to reporting procedures and 

standards may mean that the system of tolerance levels needs to be revised. 

An important question that must be considered is whether Scotland may be 

considered representative of the UK as a whole in terms of accident reporting, 

so that the conversion factors derived from Scottish data may be generalised to 

the rest of the country. The UK Department for Transport is currently carrying 

out a similar study to match English STATS19 records to in-patient data from 

the Hospital Episodes System, the equivalent for England of the SHIPS system 

in Scotland, so a well-founded answer may well emerge in due course. 

For the present, one may observe that Scotland is typical of the UK in terms of 

traffic law and traffic conditions, and the police in Scotland are subject to the 

same operational pressures as the police in the rest of the country. On the other 

hand, although Scotland does contain major urban areas, overall it contains a 




higher proportion of rural and remote areas than do England and Wales. In 

terms of hospital admission and clinical procedures, Scottish hospitals operate 

within the framework of the National Health Service, although budgets are 

devolved and this may lead to some variation across the country. 

Scotland accounts for about 9% of the UK’s fatal casualties and 6% of all 

casualties, and the study has matched data from 9 years, so there is no reason 

on statistical grounds that the results should not be considered representative. 

Overall, one may tentatively argue that the conversion factors derived from 

Scottish data may be generalised to the rest of the country. 


Apollo Project (2006). 

http://www.unav.es/preventiva/traffic_accidents/pagina_5.html 

University of Navarra – Apollo Project. 

Association for the Advancement of Automotive Medicine (1998). The 

abbreviated injury scale, 1998 revision. Des Plaines, USA. 

Broughton J, Keigan M and James F J (2001). Linkage of hospital trauma 

data and road accident data. TRL Report 518. Wokingham: TRL Limited. 

CODES. Reports available from 

http://www-nrd.nhtsa.dot.gov/departments/nrd-30/ncsa/CODES.html 

Keigan M, Broughton J and Tunbridge R J (1999). Linkage of STATS19 and 

Scottish hospital in-patient data – analyses for 1980-1995. TRL Report 420. 

Wokingham: TRL Limited. 

Nicholl JP. The Use of Hospital In-patient Data in the Analysis of the Injuries 

Sustained by Road Accident Casualties. Supplementary Report 628, 

Wokingham: TRL Limited. 

Simpson H F (1996). Comparison of hospital and police casualty data: a national 

study. TRL Report 173: Wokingham: TRL Limited. 

Stone R D (1984). Computer linkage of transport and health data. Laboratory 

Report LR1130, Wokingham: TRL Limited. 

World Health Organisation (1992). International Statistical Classification of 

Diseases and Related Health Problems. Tenth Revision. Geneva.

Estimation Real Number of Road Accident Casualties - European ...

Transform your PDFs into Flipbooks and boost your revenue!

Estimation Real Number of Road Accident Casualties - European ...

Transform your PDFs into Flipbooks and boost your revenue!

Delete template?

Save as template ?